Robust Safety-Critical Control for Dynamic Robotics
11 Optimal Robust Safety-Critical Control forDynamic Robotics
Quan Nguyen and Koushil Sreenath
Abstract —We present a novel method of optimal robust controlthrough quadratic programs that offers tracking stability whilesubject to input and state-based constraints as well as safety-critical constraints for nonlinear dynamical robotic systemsin the presence of model uncertainty. The proposed methodformulates robust control Lyapunov and barrier functions toprovide guarantees of stability and safety in the presence of modeluncertainty. We evaluate our proposed control design on dynamicwalking of a five-link planar bipedal robot subject to contactforce constraints as well as safety-critical precise foot placementson stepping stones, all while subject to model uncertainty. Weconduct preliminary experimental validation of the proposedcontroller on a rectilinear spring-cart system under differenttypes of model uncertainty and perturbations.
I. I
NTRODUCTION
Designing controllers for complex robotic systems withnonlinear and hybrid dynamics for achieving stable high-speed tracking of reference trajectories while simultaneouslyguaranteeing input, state, and safety-critical constraints ischallenging. Constraints on robotic systems arise as limits ofthe physical hardware (such as work-space constraints, jointposition and velocity constraints, and motor torque constraints)as well as constraints imposed by controllers for safe operationof the system (such as collision constraints, range constraints,connectivity constraints, contact force constraints, etc.) Fur-ther, adding to the challenges of stable constrained control isthe presence of high-levels of uncertainty in the dynamicalmodel of the robot and its interaction with the environment.The goal of this paper is to address the problem of designingstabilizing controllers for systems with strict constraints andwith model uncertainty.
A. Background
Lyapunov functions and control Lyapunov functions (CLFs)are a classical tool for design and analysis of feedback controlthat stabilize the closed-loop dynamics of both linear andnonlinear dynamical systems, see [15]. Traditional CLF-basedcontrollers involve closed-form control expressions such asthe min-norm and the Sontag controllers [43]. Recently, anovel approach of expressing CLF-based controllers via onlinequadratic programs (QPs) in [17] opened an effective way fordealing with stability while also enabling the incorporation ofadditional constraints, such as input constraints. In [17], the
Q. Nguyen is with the Department of Aerospace and Mechanical Engi-neering, University of Southern California, Los Angeles, CA 90007, email: [email protected] .K. Sreenath is with the Department of Mechanical Engineering, Universityof California, Berkeley, CA 94720, email: [email protected] . CLF-based controller is expressed in the QP as an inequalityconstraint on the time-derivative of the CLF, which easilyenables adding additional constraints such as input saturationthrough the relaxation of the CLF inequality condition. TheQP can be executed online at 1kHz in real-time as a state-dependent feedback controller. However, this work did nothandle model uncertainty or safety-critical constraints.In addition to CLFs, we also draw inspiration from recentmethods of control barrier functions (CBFs) that can beincorporated with control Lyapunov function based quadraticprograms to result in the CBF-CLF-QP, as introduced in [1].This framework enables handling safety-critical constraintseffectively in real-time. Experimental validation of this typeof controller for the problem of Adaptive Cruise Controlwas presented in [32]. This framework has also been ex-tended to various interesting application domains, such assafety-critical geometric control for quadrotor systems [51]and safety-critical dynamic walking for bipedal robots [37],[22]. Although this work can handle safety-critical constraints,however a precise model of the system is required to enforcethe constraints.Moreover, as presented in [52], preliminary robustnessanalysis of the CBFs indicate that the safety-critical constraintwill be violated in the presence of model uncertainty, withthe amount of violation being bounded by the value of theupper bound of the model uncertainty. In particular, modeluncertainty leads to constraint violation of the safety-criticalconstraints. Recent efforts to improve the robustness of CBFsusing Gaussian Process learning were introduced in [13] and[7]. In this paper, we seek a method to simultaneously handlerobust stability, robust input-based constraints and robust state-dependent constraint in the presence of model uncertainty. Wewill do this through robust control formulations of the CLF,CBF, and constraints.The framework can be easily achieved byextending the nominal CBF-CLF-QP controller while improv-ing significantly the robustness of the entire approach with aprovable stability and safety guarantee.QP based controllers for robotic systems have been increas-ingly used in recent years. Especially, the DARPA RoboticsChallenge inspired several new methods in this direction [14],[9], wherein QP controllers are used to formulate inverse kine-matic and inverse dynamic control problems while minimizingtracking of desired accelerations. While these QP controllersenforce tracking performance and constraints at each time step,our method uses CLF and CBF condition to guarantee stabilityand safety constraints for future time.For control of constrained systems, Model Predictive Con-trol (MPC) has been widely used in many industrial applica- a r X i v : . [ ee ss . S Y ] M a y tions [40]. However, the method is computationally expensive.Even with a linear model and polyhedral constraints, thetraditional approach can require a solving time from secondsup to minutes. Recent advanced approach in [49] attempts toreduce the computation time but still needs at least tens ofmilliseconds for such systems. Furthermore, MPC just checksthose polyhedral constraints for a finite time or horizon (often10 to 30 sampling times). The CBF approach instead enforcesnonlinear state-dependent constraints of a nonlinear system.The condition is also guaranteed for future time or in otherwords for an infinite horizon by enforcing only one scalarcondition that can be solved in 1 ms.Robust control is an extensively studied topic. We haveestablished methods, such as H ∞ -based robust control andlinear quadratic Guassian (LQG) based robust control [27, 53]for robust control of linear systems. For robust control ofnonlinear systems, input-to-state stability (ISS) and slidingmode control (SMC) are two main methods. The ISS technique(see [44, 45, 3]) can be used to both analyze the robustnessof nonlinear systems as well as design robust controllersbased on control Lyapunov functions. However, a primarydisadvantage of ISS based controllers is that the resulting con-troller only maintains the system errors in a sufficiently smallneighborhood of the origin, resulting in non-zero trackingerrors. In recent years, there has been work on robust controlof hybrid systems based on the ISS technique, for instancesee [48, 21, 6]. In contrast, sliding mode control techniquescan deal with a wide range of uncertainties and drive thesystem to follow a given sliding surface, thereby drivingoutputs to desired values without any tracking errors (see[11, 10, 47, 12]). However, the primary disadvantage of SMCis the chattering phenomenon caused by discrete switching forkeeping the system attracted to the sliding surface.Robust control techniques have also been extensively ap-plied to robotic manipulator arms, see [46, 29], however ma-nipulator arms do not have challenges such as underactuationand unilateral ground contact forces making their controleasier. Robust control techniques have also been appliedto bipedal walking. For instance, the work in [28] extendsadaptive robust control of manipulator arms to bipedal robots,the work in [24] considers a simple 2D inverted-pendulummodel and pre-computes a control policy through offlinenonlinear optimization to prevent falls under the assumptionof bounded disturbances. Similarly, recent results on robustfeedback motion planning for the problem of UAVs avoidingobstacles in [30] also precomputed a library of “funnels” viaconvex optimization that represents different maneuvers of thesystem under bounded disturbances. A real-time planner thencomposes motion trajectories based on the resulting funnellibrary. Recent work in reachability-based analysis has beenextended to enforce safety-constraints with model uncertaintyby safe trajectory synthesis [26, 16]. Offline optimizationfor stabilization of walking and running with robustness todiscrete-time uncertainties such as terrain perturbations hasbeen carried out in [19, 20]. Our method, in contrast, isbased on real-time feedback controller to guarantee robuststability as well as robust safety of robotic systems through anonline optimization without the need of precomputed motion plans. Additional robust planning and control techniques existfor legged robots where the robustness is with respect tostochastic uncertainty in the model / terrain, for instance see[23, 42, 4, 5]. B. Contributions
The main contributions of this paper with respect to priorwork are as follows: • Introduction of a new technique that optimally introducesrobust control via quadratic programs to handle stability,input-based constraints and state-dependent constraintsunder high levels of model uncertainty. • Robust stability and robust safety-critical constraintsachieved through a min-max inequality constraint onthe time-derivative of a control Lyapunov function andcontrol barrier function respectively. • Theoretical stability analysis for the QP controllers withrelaxed CLF inequality. • Numerical validation of the proposed controller on dif-ferent problems: – Dynamic walking of a bipedal robot while carryingan unknown load and subject to contact force con-straints; and – Dynamic walking of a bipedal robot while carryingan unknown load, subject to contact force constraintsand precise foot-step location constraints. • Experimental validation of the proposed control methodon a rectilinear spring-cart system.Note that preliminary results of this work were presentedin [36, 39]. In contrast to the preliminary results, this paperpresents (a) an entirely new min-max formulation of theproposed controller, (b) additional numerical examples forcontrol of a bipedal robot, (c) experimental validations of theproposed controller on a rectilinear spring-cart system and (d)detailed stability and safety analysis for QP controllers withrelaxed CLF inequality and the proposed robust CBF-CLF-QPcontroller.
C. Organization
The rest of the paper is organized as follows. SectionII revisits control barrier functions and control Lyapunovfunctions based quadratic programs (CBF-CLF-QPs). SectionIII presents the stability analysis for the QP controllers withrelaxed CLF inequality. Section IV discusses the adverseeffects of uncertainty and then presents the proposed optimalrobust control and formulates as a CBF-CLF-QP. Section Vpresents numerical and experimental validation on differentdynamical robotic systems. Finally, Section VI provides con-cluding remarks.
II. C
ONTROL L YAPUNOV F UNCTIONS AND C ONTROL B ARRIER F UNCTION BASED Q UADRATIC P ROGRAMS R EVISITED
A. Model and Input-Output Linearizing Control
Consider a nonlinear control affine hybrid model H : (cid:40) ˙ x = f ( x ) + g ( x ) u, x / ∈ Sx + = ∆( x − ) , x ∈ S (1) y = y ( x ) , where x ∈ R n is the system state, u ∈ R m is the control input, S is the switching surface of the hybrid system, and y ∈ R m is a set of outputs.If the control output y ( x ) has relative degree 2, then thetime-derivative ˙ y ( x ) will be a function of the state x and notdependent on the control input u . Considering the second time-derivative ¨ y , we have: ¨ y = ∂ ˙ y∂x ˙ x = L f y ( x ) + L g L f y ( x ) u, (2)where L represents the Lie derivatives. To be more specific: L f y ( x ) (cid:44) ∂ ˙ y∂x f ( x ) , L g L f y ( x ) (cid:44) ∂ ˙ y∂x g ( x ) . (3)If the decoupling matrix L g L f y ( x ) is invertible, then thecontroller u ( x, µ ) = u ff ( x ) + ( L g L f y ( x )) − µ, (4)with the feed-forward control input u ff ( x ) = − ( L g L f y ( x )) − L f y ( x ) , (5)input-output linearizes the system. The dynamics of the sys-tem (1) can then be described in terms of dynamics of thetransverse variables, η ∈ R m , and the coordinates z ∈ Z with Z being the co-dimension m manifold Z = { x ∈ R n | η ( x ) ≡ } . (6)One choice for the transverse variables is, η = (cid:20) y ( x )˙ y ( x ) (cid:21) . (7)The input-output linearized hybrid system then is, H IO : ˙ η = ¯ f ( η ) + ¯ g ( η ) µ, ˙ z = p ( η, z ) , ( η, z ) / ∈ S,η + = ∆ η ( η − , z − ) ,z + = ∆ Z ( η − , z − ) , ( η − , z − ) ∈ S. (8) y = y ( η ) , where z represents uncontrolled states [2], and ¯ f ( η ) = F η, ¯ g ( η ) = G, (9)with, F = (cid:20) O IO O (cid:21) and G = (cid:20) OI (cid:21) . (10)The linear system in (9) is in controllable canonical form,and a linear controller such as µ = − Kη can be designedsuch that the closed-loop system ˙ η = ( F − GK ) η is stable.A corresponding quadratic Lyapunov function can then beestablished through the Lyapunov equation. B. Control Lyapunov Function based Quadratic Programs1) CLF-QP:
Instead of a linear control design µ = − Kη in(4), an alternative control design is through a control Lyapunovfunction V ( η ) , wherein a control is chosen pointwise intime such that the time deriviative of the Lyapunov function ˙ V ( η, µ ) ≤ , resulting in stability in the sense of Lyapunov,or ˙ V ( η, µ ) < for asymptotic stability, or ˙ V ( η, µ )+ λV ( η ) ≤ , λ > for exponential stability.To enable directly controlling the rate of convergence,we use a rapidly exponentially stabilizing control Lyapunovfunction (RES-CLF) , introduced in [2]. RES-CLFs provideguarantees of rapid exponential stability for the transversevariables η . In particular, a function V ε ( η ) is a RES-CLF forthe system (1) if there exist positive constants c , c , c > such that for all < ε < and all states ( η, z ) it holds that c (cid:107) η (cid:107) ≤ V ε ( η ) ≤ c ε (cid:107) η (cid:107) , (11) ˙ V ε ( η, µ ) + c ε V ε ( η ) ≤ . (12)The RES-CLF will take the form: V ε ( η ) = η T (cid:20) ε I I (cid:21) P (cid:20) ε I I (cid:21) η =: η T P ε η, (13)and the time derivative of the RES-CLF (13) is computed as ˙ V ε ( η, µ ) = ∂V ε ∂η ˙ η = L ¯ f V ε ( η ) + L ¯ g V ε ( η ) µ, (14)where L ¯ f V ε ( η ) = ∂V ε ∂η ¯ f ( η ) = η T ( F T P ε + P ε F ) η,L ¯ g V ε ( η ) = ∂V ε ∂η ¯ g ( η ) = 2 η T P ε G. (15)It can be show that for any Lipschitz continuous feedbackcontrol law µ that satisfies the RES condition (12), it holdsthat V ( η ) ≤ e − c ε t V ( η (0)) , (cid:107) η ( t ) (cid:107) ≤ ε (cid:114) c c e − c ε t (cid:107) η (0) (cid:107) , (16)i.e. the rate of exponential convergence can be directly con-trolled with the constant ε through c ε . One such controller isthe CLF-based quadratic program (CLF-QP)-based controller,introduced in [17], where µ is directly selected through anonline quadratic program to satisfy (12): CLF-QP : u ∗ ( x ) = argmin u,µ µ T µ (17)s.t. ˙ V ε ( η, µ ) + c ε V ε ( η ) ≤ , ( CLF ) u = u ff ( x ) + ( L g L f y ( x )) − µ. ( IO )Note that the above minimization problem is a quadraticprogram since the ineqality constraint on the time-derivativeof the Lyapunov function can be written as a linear ineqalityconstraint A clf µ ≤ b clf , (18) where A clf = L ¯ g V ε ( η ); b clf = − L ¯ f V ε ( η ) − c ε V ε ( η ) . (19)Moreover, the IO equality constraint is linear in u, µ , and canbe written as: A IO (cid:20) uµ (cid:21) = b IO , (20)where A IO = (cid:2) I − ( L g L f y ( x )) − (cid:3) , b IO = u ff ( x ) . (21)This optimization is solved pointwise in time. Furthermorethe existence of efficient quadratic program solvers, such asCVXGEN [31], enable solving the problem in real-time speedsover kHz. Remark 1:
Note that the solution of the above QP has beenshown to be Lipschitz continuous in [34].
2) CLF-QP with Constraints:
Formulating the controlproblem as a quadratic program now enables us to incorporateconstraints into the optimization. These constraints could beinput constraints for input saturation or state-based constraintssuch as friction constraints, contact force constraints, etc.,for robotic locomotion and manipulation. These types ofconstraints can be expressed in a general form as A c ( x ) u ≤ b c ( x ) . (22)The CLF-QP based controller with additional constraintsthen takes the form, CLF-QP with Constraints : u ∗ ( x ) =argmin u,µ,δ µ T µ + pδ (23)s.t. ˙ V ε ( η, µ ) + c ε V ε ( η ) ≤ δ, ( CLF ) A c ( x ) u ≤ b c ( x ) , ( Constraints ) u = u ff ( x ) + ( L g L f y ( x )) − µ, ( IO )where p is a large positive number that represents the penaltyof relaxing the CLF inequality, which is necessary to keep theQP feasible when the constraints conflict with each other.The constraints above could be input saturation constraintsexpressed as, u ∗ ( x ) =argmin u,µ,δ µ T µ + pδ (24)s.t. ˙ V ε ( η, µ ) + c ε V ε ( η ) ≤ δ, ( CLF ) u min ≤ u ≤ u max , ( Input Saturation ) u = u ff ( x ) + ( L g L f y ( x )) − µ. ( IO )Note that, similar to (18), the CLF inequality condition inthe above CLF-QPs is affine in µ , ensuring that these areactually quadratic programs.This formulation opened a novel method to guarantee stabil-ity of nonlinear systems with respect to additional constraintssuch as torque saturation [17], wherein experimental demon-stration of bipedal walking with strict input constraints was demonstrated, and in enabling the application of L1 adaptivecontrol for bipedal robots [35]. Remark 2: (Stability of CLF-QP with relaxed CLF in-equality.)
In subsequent sections of this paper, we developdifferent types of controllers based on the relaxed CLF-QP(23). Allowing the violation of the RES-CLF condition (12)enables us to incorporate various other input and state con-straints as we saw in (23). Additionally, the relaxed RES-CLFcondition also enables incorporating safety constraints throughbarriers (see Section II-C), as well as enabling modificationof the controller to increase the robustness of the closed-loopsystem (see Section IV). However, it must be noted that therelaxation of the RES-CLF condition could lead to potentialinstability. In Section III, we establish sufficient conditionsunder which the relaxed CLF-QP controller can still retain theexponential stability of the hybrid periodic orbit.Having revisited control Lyapunov function based quadraticprograms, we will next revisit control Barrier functions.
C. Control Barrier Function
We begin with the control affine system (1) with the goalto design a controller to keep the state x in the safe set C = { x ∈ R n : h ( x ) ≥ } , (25)where h : R n → R is a continuously differentiable function.Then a function B : C → R is a Control Barrier Function(CBF) [1] if there exists class K function α and α suchthat, for all x ∈ Int ( C ) = { x ∈ R n : h ( x ) > } , α ( h ( x )) ≤ B ( x ) ≤ α ( h ( x )) , (26) ˙ B ( x, u ) ≤ γB ( x ) , (27)where ˙ B ( x, u ) = ∂B∂x ˙ x = L f B ( x ) + L g B ( x ) u, (28)with the Lie derivatives computed as, L f B ( x ) = ∂B∂x f ( x ) , L g B ( x ) = ∂B∂x g ( x ) . (29)Thus, if there exists a Barrier function B ( x ) that satisfies theCBF condition in (27), then C is forward invariant, or in otherwords, if x (0) = x ∈ C , i.e., h ( x ) ≥ , then x = x ( t ) ∈C , ∀ t , i.e., h ( x ( t )) ≥ , ∀ t . Note that, as mentioned in [1], thisnotion of a CBF is stricter than standard notions of CBFs inprior literature that only require ˙ B ≤ .In this paper, we will use the following reciprocal controlBarrier candidate function: B ( x ) = 1 h ( x ) . (30)Incorporating the CBF condition (27) into the CLF-QP, we have the following CBF-CLF-QP based controllers: CBF-CLF-QP : u ∗ ( x ) =argmin u,µ,δ µ T µ + p δ (31)s.t. ˙ V ε ( η, µ ) + c ε V ε ( η ) ≤ δ, ( CLF ) ˙ B ( x, u ) − γB ( x ) ≤ , ( CBF ) A c ( x ) u ≤ b c ( x ) , ( Constraints ) u = u ff ( x ) + ( L g L f y ( x )) − µ. ( IO ) Remark 3:
Note that, similar to (18), for any nonlinearaffine system and nonlinear state-dependent constraint, theCBF condition (27) is scalar and affine in u , ensuring acompact quadratic program that can be solved in 1 kHz. Moreimportantly, the satisfaction of this instantaneous condition ateach time guarantees the state dependent constraint h ( x ) ≥ for all t ≥ t .Having presented control Lyapunov functions, control Bar-rier functions, and their incorporation into a quadratic programwith constraints, we now prove the stability of the QP con-troller with relaxed CLF inequality.III. S UFFICIENT CONDITIONS FOR THE STABILITY OF
CLF
WITH RELAXED INEQUALITY
In this Section, we will present two theorems and their proofabout the stability of CLF based controller with relaxed RES-CLF condition for both continuous-time and hybrid systems.We begin with the standard RES-CLF that guarantees thefollowing inequality, ˙ V ε ( η, µ ) + c ε V ε ( η ) ≤ . (32)The CLF-QP with the relaxed inequality takes the form, ˙ V ε ( η, µ ) + c ε V ε ( η ) ≤ d ε , (33)where d ε ( t ) ≥ represents the time-varying relaxation of theRES-CLF condition. We define w ε ( t ) = (cid:90) t d ε ( τ ) V ε d τ. (34)to represent a scaled version of the total relaxation up to time t . In the following subsections, we make use of this quantityto establish exponential stability under certain conditions forboth continuous-time and hybrid systems. In Section III-Awe will look at continuous-time systems where we will needexponential stability ( ε = 1 in the above formulations), whilein Section III-B we will look at hybrid systems where we needrapid exponential stability ( ε < ). A. Stability of the Relaxed CLF-QP Controller forContinuous-Time Systems
Consider a control affine system ˙ η = ¯ f ( η ) + ¯ g ( η ) µ, (35) ˙ z = p ( η, z ) , where η is the controlled (or output) state, and z is theuncontrolled state. Let O Z be an exponentially stable periodicorbit for the zero dynamics ˙ z = p (0 , z ) . Let O = ι ( O Z ) be aperiodic orbit for the full-order dynamics corresponding to theperiodic orbit of the zero dynamics, O Z , through the canonicalembedding ι : Z → X × Z given by ι ( z ) = (0 , z ) . As statedin [2, Theorem 1], for all control inputs µ ( η, z ) that guaranteeenforcing the ES-CLF condition (with ε = 1 ) (32), we havethat O = ι ( O Z ) is an exponentially stable periodic orbit of(35). Then, for the relaxed CLF-QP condition (33) ( d ε (cid:54)≡ ),the following theorem establishes sufficient conditions underwhich the exponential stability of the periodic orbit still holds. Theorem 1:
Consider a nonlinear control affine system(35). Let O Z be an exponentially stable periodic orbit forthe zero dynamics ˙ z = p (0 , z ) . Then O = ι ( O Z ) is anexponentially stable periodic orbit of (35), if ¯ w ε := sup t ≥ w ε ( t ) (36)is a finite number. Proof:
Note that if d ε ( t ) ≡ , i.e., there is no violation onthe RES-CLF condition, we will have conventional exponentialstability as stated in [2]. Here, we will extend the proof ofexponential stability to the case of relaxation of the controlLyapunov function condition when ¯ w ε is finite.We begin by noting that since d ε ( t ) ≥ , we have w ε ( t ) ≤ ¯ w ε , ∀ t ≥ . (37)Next, from (33), we have dV ε dt ≤ − c ε V ε + d ε ( t ) , ⇒ dV ε V ε ≤ − c ε dt + d ε ( t ) V ε dt, ⇒ ln (cid:18) V ε ( t ) V ε (0) (cid:19) ≤ − c ε t + (cid:90) t d ε ( τ ) V ε d τ, ⇒ V ε ( t ) ≤ e − c ε t + w ε ( t ) V ε (0) . (38)This, in combination with the inequality in (11), thenimplies that (cid:107) η ( t ) (cid:107) ≤ (cid:114) c c ε e − c ε t + w ε ( t ) (cid:107) η (0) (cid:107) . (39)Moreover, from the inequality in (37), we have (cid:107) η ( t ) (cid:107) ≤ (cid:114) c c ε e − c ε t + ¯ w ε (cid:107) η (0) (cid:107) , = (cid:18)(cid:114) c c ε e ¯ w ε (cid:19) e − c ε t (cid:107) η (0) (cid:107) . (40)Therefore, if ¯ w ε is finite, the control output η will still beexponentially stable under the relaxed CLF condition (33).Therefore, from [2, Theorem 1], O = ι ( O Z ) is an exponen-tially stable periodic orbit of (35). Remark 4:
It must be noted that the CLF V ε no longer pro-vides a guarantee of exponential stability due to the relaxationin (33). However due to the inequality established in (40), andby the converse Lyapunov function theorems [25], there existsanother CLF ˜ V ε that satisfies ˜ c || η || ≤ ˜ V ε ( η ) ≤ ˜ c ε || η || , and ˙˜ V ε ( η, µ ) + ˜ c ε ˜ V ε ( η ) ≤ , for some positive constants ˜ c , ˜ c , ˜ c ,and guarantees exponential stability.In the next section, we consider the hybrid system withimpulse effects, wherein we will need to take into account theimpact time or the switching time that signifies the end of thecontinuous-time phase and involves a discrete-time jump inthe state. B. Stability of the Relaxed CLF-QP Controller for HybridSystems
Here we look at the stability of the relaxed CLF-QPcontroller for hybrid systems of the form as defined in (8)without the restriction on the vector fields as in (9). Similarto the stability analysis for continuous-time systems in theprevious section, we also use here the notions of d ε ( t ) (33),the relaxation of the CLF condition, and w ε ( t ) (34), the scaledversion of the total relaxation up to time t .We also define T εI ( η, z ) to be the time-to-impact or timetaken to go from the state ( η, z ) to the switching surface S . Then, intuitively, w ε ( T εI ( η, z )) indicates a scaled versionof the total violation of the RES-CLF bound in (32) overone complete step. If w ε ( T εI ( η, z )) ≤ , it implies that V ε ( T εI ( η, z )) is less than or equal to what would have resultedif the RES-CLF bound had not been violated at all. As wewill see in the following theorem, we will in fact only require w ε ( T εI ( η, z )) to be upper bounded by a positive constant forexponential stability.We first define the hybrid zero dynamics as the hybriddynamics (8) restricted to the surface Z in (6), i.e., H Z : (cid:40) ˙ z = p (0 , z ) , z / ∈ S ∩ Z,z + = ∆ z (0 , z − ) , z − ∈ S ∩ Z. (41)We then have the following theorem: Theorem 2:
Let O Z be an exponentially stable hybrid pe-riodic orbit of the hybrid zero dynamics H | Z (41) transverseto S ∩ Z and the continuous dynamics of H (8) controlled bya CLF-QP with relaxed inequality (23) . Then there exists an ε > and an ¯ w ε ≥ such that for each < ε < ε , if the solu-tion µ ε ( η, z ) of the CLF-QP (23) satisfies w ε ( T εI ( η, z )) ≤ ¯ w ε ,then O = ι ( O Z ) is an exponentially stable hybrid periodicorbit of H .Proof: See Appendix A in the supplementary document.Having presented the stability analysis for the QP controllerwith relaxed CLF inequality, in the next Section, we willexplore the effects of model uncertainty and propose robustQP controllers to guarantee stability, safety and constraintsunder model uncertainty.IV. P
ROPOSED R OBUST C ONTROL BASED Q UADRATIC P ROGRAMS
The controllers presented in Section II provide means of anoptimal control scheme that balances conflicting requirementsbetween stability, state-based constraints and energy consump-tion. It is a powerful method that has been deployed success-fully for different applications, for example Adaptive Cruise
Notations Model types f, g true (unknown) nonlinear model ˜ f , ˜ g nominal nonlinear model ¯ f , ¯ g true (unknown) I-O linearized model ˜¯ f, ˜¯ g nominal I-O linearized model TABLE I: A list of notations for different models used inthis paper. A true model represents the actual (possibly notperfectly known) model of the physical system, while thenominal model represents the model that the controller uses.While most controllers assume the true model is known, therobust controllers in this paper use the nominal models andoffer robustness guarantees to the uncertainty between the twomodels.Control [32], quadrotor flight [51], and dynamic walking forbipedal robots [37], [22].However, a primary disadvantage of the CBF-CLF-QPcontroller is that it requires the knowledge of an accuratedynamical model of the plant to make guarantees on stabilityas well as safety. In particular, enforcing the CLF and CBFconditions in the QP (31) requires the accurate knowledgeof the nonlinear dynamics f ( x ) , g ( x ) to compute the Liederivatives in (3), (15) and (29). Therefore, model uncertaintythat is usually present in physical systems can potentiallycause poor quality of control leading to tracking errors, andlead to instability [36], as well as violation of the safety-critical constraints [52]. In this section, we will explore theeffect of uncertainty on the CLF-QP controller, input and stateconstraints, and safety constraints enforced by the CBF-QPcontroller and then proposed robust controllers based quadraticprograms to address those effects. Firstly, we will analyze theeffects of model uncertainty on the CLF condition. A. Adverse Effects of Uncertainty on the CLF-QP controller
In order to analyze the effects of model uncertainty in ourcontrollers, we assume that the vector fields, f ( x ) , g ( x ) of thereal dynamics (1), are unknown. We therefore have to designour controller based on the nominal vector fields ˜ f ( x ) , ˜ g ( x ) .Then, the pre-control law (4) get’s reformulated as u ( x ) = ˜ u ff ( x ) + ( L ˜ g L ˜ f y ( x )) − µ, (42)with ˜ u ff ( x ) := − ( L ˜ g L ˜ f y ( x )) − L f y ( x ) , (43)where we have used the nominal model rather than theunknown real dynamics.Substituting u ( x ) from (42) into (2), the input-output lin-earized system then becomes ¨ y = µ + ∆ + ∆ µ, (44)where ∆ ∈ R m , ∆ ∈ R m × m , are given by, ∆ := L f h ( x ) − L g L f h ( x )( L ˜ g L ˜ f h ( x )) − L f h ( x ) , ∆ := L g L f h ( x )( L ˜ g L ˜ f h ( x )) − − I. (45) Remark 5:
In the definitions of ∆ , ∆ , note that whenthere is no model uncertainty, i.e., ˜ f = f, ˜ g = g , then ∆ =∆ = 0 .Using F and G as in (10), the closed-loop system now takesthe form ˙ η = ˜¯ f ( η ) + ˜¯ g ( η ) µ, (46)where ˜¯ f ( η ) = F η + (cid:20) O ∆ (cid:21) , ˜¯ g ( η ) = G + (cid:20) O ∆ (cid:21) . (47)In fact for ∆ (cid:54) = 0 , the closed-loop system does not have anequilibrium, and for ∆ (cid:54) = 0 , the controller could potentiallydestabilize the system (see [41]). This raises the question ofwhether it’s possible for controllers to account for this modeluncertainty, and if so, how do we design such a controller.In particular, the time-derivative of the CLF in (14) becomesmore complex and dependent on ∆ , ∆ .Again, because our controller can only use the nominalmodel ˜¯ f, ˜¯ g in (46) and not the true model ¯ f , ¯ g as defined in (9)(see Table I for different types of models), the Lie derivativesof V ε with respect to the nominal model ˜¯ f, ˜¯ g will be as follows: L ˜¯ f V ε = η T ( F T P ε + P ε F ) η, (48) L ˜¯ g V ε = 2 η T P ε G. (49)Then, the true time-derivative of the CLF defined in (13) is: ˙ V ε = L ¯ f V ε + L ¯ g V ε µ = L ˜¯ f V ε + 2 η T P ε (cid:20) ∆ (cid:21)(cid:124) (cid:123)(cid:122) (cid:125) ∆ v + L ˜¯ g V ε µ (cid:124) (cid:123)(cid:122) (cid:125) µ v + L ˜¯ g V ε ∆ µ (cid:124) (cid:123)(cid:122) (cid:125) ∆ v µ v = L ˜¯ f V ε + ∆ v + (1 + ∆ v ) µ v , (50)where we have defined the following new scalar variables:uncertainty ∆ v ∈ R , virtual control input µ v ∈ R , anduncertainty ∆ v ∈ R . B. Robust CLF-QP
Having discussed the effect of model uncertainty on thecontrol Lyapunov function based controllers, we now developa robust controller that can guarantee tracking and stabilityin the presence of bounded uncertainty. As we will see, bothstability and tracking performance (rate of convergence anderrors going to zero) are still retained for all uncertainty withina particular bound. For uncertainty that exceeds the specifiedbound, there is graceful degradation in performance.With the presence of model uncertainty, the CLF condition(12) now becomes: ˙ V ε ( η, ∆ v , ∆ v , µ v ) + c ε V ε ≤ . (51)In general, we can not satisfy this inequality for all possibleunknown ∆ v , ∆ v . To address this, we assume the uncertaintyis bounded as follows | ∆ v | ≤ ∆ v ,max , | ∆ v | ≤ ∆ v ,max . (52) Under this assumption, we have the following robust CLFcondition: ˙ V ε ( η, ∆ v , ∆ v , µ v ) + c ε V ε ≤ , ∀| ∆ v | ≤ ∆ v ,max , ∀| ∆ v | ≤ ∆ v ,max ⇔ max | ∆ v |≤ ∆ v ,max | ∆ |≤ ∆ v ,max ˙ V ε ( η, ∆ v , ∆ v , µ v ) + c ε V ε ≤ . (53)It means that choosing µ that satisfies (53) implies (16)for every ∆ v , ∆ v satisfying | ∆ v | ≤ ∆ v ,max , | ∆ v | ≤ ∆ v ,max . Then, our optimal robust control canbe expressed as the following min-max problem: Robust CLF : u ∗ ( x ) =argmin u,µ,µ v µ T µ (54)s.t. max | ∆ v |≤ ∆ v ,max | ∆ v |≤ ∆ v ,max ˙ V ε ( η, ∆ v , ∆ v , µ v ) + c ε V ε ≤ µ v = L ˜¯ g V ε µ ( Robust CLF ) u = ˜ u ff ( x ) + ( L ˜ g L ˜ f y ( x )) − µ. ( IO )Note that, one advantage of CLF based control is to convertthe problem of driving a vector output η → to the problemof driving a scalar V ε ( η ) → . Therefore, since V ε is a scalar,we always have µ v , ∆ v , ∆ v in (50) being scalars. Then themin-max optimization problem (54) can be expressed as aquadratic program as follows: max | ∆ v |≤ ∆ v ,max | ∆ |≤ ∆ v ,max ˙ V ε ( η, ∆ v , ∆ v , µ v ) + c ε V ε ≤ ⇔ max | ∆ v |≤ ∆ v ,max | ∆ |≤ ∆ v ,max L ˜¯ f V ε + ∆ v + (1 + ∆ v ) µ v + c ε V ε ≤ , (55)and the above robust CLF condition holds if (cid:40) Ψ max + Ψ p µ v ≤ max + Ψ n µ v ≤ , (56)where Ψ max = L ˜¯ f V + c ε V ε + ∆ v ,max Ψ p = 1 + ∆ v ,max , Ψ n = 1 − ∆ v ,max . (57)Thus the robust optimal CLF in (54) can be satisfiedusing the following quadratic program based controller: Robust CLF-QP : u ∗ =argmin u,µ,µ v µ T µ (58)s.t. Ψ max ( η, ∆ v ,max ) + Ψ p (∆ v ,max ) µ v ≤ , Ψ max ( η, ∆ v ,max ) + Ψ n (∆ v ,max ) µ v ≤ ,µ v = L ˜¯ g V ε µ, ( Robust CLF ) u = ˜ u ff ( x ) + ( L ˜ g L ˜ f y ( x )) − µ. ( IO ) Remark 6:
Note that since V ( η ) is a local CLF for thenonlinear system (1) and the fact that ˙ V depends continuouslyon ∆ v , ∆ v , there always exists µ satisfying (53) for a localregion around η = 0 , ∆ v = 0 , ∆ v = 0 . Moreover, if ∆ v ,max , ∆ v ,max are chosen to be within this region, then theabove QP is guaranteed to be feasible. Remark 7:
Note that ∆ v ,max , ∆ v ,max are design parame-ters for control design. While we do not present a principledway of choosing these parameters, it must be noted that ∆ v ,max , ∆ v ,max can be chosen based on knowledge of theexpected uncertainty the controller is to be designed to handle.Having developed the robust version of the CLF-QP basedcontroller, we now incorporate constraints into the robustcontrol formulation. We note that this first formulation is fornon-robust constraints, i.e., these constraints are only evaluatedon the nominal model available to the controller. Thus thiscontrol is valid and makes sense for only those constraintsthat are not dependent on the true model.The incorporation of constraints into the CLF-QP con-troller required relaxation of the CLF condition (see [18]).Similarly, here we relax the robust CLF condition to obtain, Robust CLF-QP with Constraints : u ∗ ( x ) = argmin u,µ,µ v ,δ ,δ µ T µ + p δ + p δ (59)s.t. Ψ max ( η, ∆ v ,max ) + Ψ p (∆ v ,max ) µ v ≤ δ , Ψ max ( η, ∆ v ,max ) + Ψ n (∆ v ,max ) µ v ≤ δ ,µ v = L ˜¯ g V ε µ, ( Robust CLF ) A c ( x ) u ≤ b c ( x ) , ( Constraints ) u = u ff ( x ) + ( L g L f y ( x )) − µ. ( IO ) Remark 8:
It’s critical to note that the additional con-straints in (59) are not robust, and that this method is onlyapplicable for constraints that are invariant to model uncer-tainty. Constraints such as torque saturation, like in (24), areinvariant to model uncertainty since the control inputs arecomputed directly from the nominal model ˜ f ( x ) , ˜ g ( x ) , and donot depend on the real model f ( x ) , g ( x ) . However, constraintsthat are invariant to the model uncertainty are not common,and we will have to explicitly address and capture the effectof uncertainty on constraints. We will address this in SectionIV-D. C. Robust CBF
Following the same approach we use for the robust CLF,we now analyze the effect of model uncertainty on the CBFcondition and then present our robust CBF controller.Similar to what we have seen about the effect of uncertaintyon CLFs, we will now see the effect of uncertainty on CBFs.We note that the time-derivative of the Barrier function in (30) depends on the real model. Therefore we need to enforce thefollowing constraint given by (27): ˙ B ( x, f, g, u ) = L f B ( x ) + L g B ( x ) u ≤ γB ( x ) (60)where ˙ x = f ( x ) + g ( x ) u is the real system dynamic. As seenin the case of control Lyapunov functions and constraints,naively enforcing this barrier constraint using the nominalmodel results in, ˙ B ( x, ˜ f , ˜ g, u ) = L ˜ f B ( x ) + L ˜ g B ( x ) u ≤ γB ( x ) (61)where ˙ x = ˜ f ( x ) + ˜ g ( x ) u is the nominal system dynamicsknown by the controller. Clearly due to model uncertainty,or the difference between ( f ( x ) , g ( x )) and ( ˜ f ( x ) , ˜ g ( x )) , theconstraint in (61) is different from the one in (60). In fact, asanalyzed in [52], this results in violation of the safety-criticalconstraint established by the Barrier function.We then define a virtual control input µ b ∈ R such that thetime derivative of the CBF is ˙ B ( x, u ) = L f B ( x ) + L g B ( x ) u =: µ b . (62)We call this virtual input-output linearization (VIOL). TheCBF now takes the form of a linear system, ˙ B = µ b , andtherefore the effect of uncertainty can be easily addressedby using the same approach as with the robust CLF-QP. Inparticular, similar to what we have done with CLF in SectionIV-A, the effect of model uncertainty on CBF transforms (60)into the following form: ˙ B ( x, ∆ b , ∆ b , µ b ) = L f B ( x ) + L g B ( x ) u = µ b + ∆ b + ∆ b µ b , (63)with ∆ b , ∆ b ∈ R and µ b is defined based on the nominalmodel ˜ f , ˜ g : µ b = L ˜ f B ( x ) + L ˜ g B ( x ) u. (64)The CBF condition (60) under model uncertainty becomes ˙ B ( x, ∆ b , ∆ b , µ b ) − γB ( x ) ≤ . (65)Because we developed the CLF and CBF having the similarform of I-O linearized system, we now have a systematic wayto develop the Robust CBF-CLF-QP. Again, we will assumethat our model uncertainty is bounded, i.e., | ∆ b | ≤ ∆ b ,max , | ∆ b | ≤ ∆ b ,max . (66)Then, we will have the robust CBF condition based on thisassumption as, max | ∆ b |≤ ∆ b ,max | ∆ b |≤ ∆ b ,max ˙ B ( x, ∆ b , ∆ b , µ b ) − γB ( x ) ≤ . (67)Note that choosing µ b that satisfies (67) implies x ∈ C or B ( x ) ≥ for every ∆ b , ∆ b satisfying | ∆ b | ≤ ∆ b ,max , | ∆ b | ≤ ∆ b ,max . Here, C is as defined in (25) and B ( x ) is as definedin (30). Note that the robust CBF condition in (67) is affinein µ b and can be expressed as max | ∆ b |≤ ∆ b ,max | ∆ b |≤ ∆ b ,max Ψ b + Ψ b µ b ≤ (68) where “ b ” refers to the CBF constraint, and Ψ b ( x, ∆ b ) := ∆ b − γB ( x ) , Ψ b ( x, ∆ b ) := 1 + ∆ b , (69)where the above arises due to the time-derivative of the CBFfrom (63). Since Ψ b , Ψ b are affine with respect to ∆ b , ∆ b ,and with the assumptions on the bounds on the uncertainty in(66), the robust CBF condition (68) will hold if the followingtwo inequalities hold Ψ b ,max ( x ) + Ψ b ,p ( x ) µ b ≤ , Ψ b ,max ( x ) + Ψ b ,n ( x ) µ b ≤ . (70)where Ψ b ,max := ∆ b ,max − γB ( x ) , (71) Ψ b ,p := 1 + ∆ b ,max , (72) Ψ b ,n := 1 − ∆ b ,max , (73)We then can incorporate the above robust CBF conditionsinto the robust CLF-QP (59) resulting in a quadratic program.Having robustified CLFs and CBFs, we will now apply thisto obtain robust constraints. D. Robust Constraints
As developed in Section II, the input and state constraintscan be expressed as (22). These constraints depend on themodel explicitly and are constraints on the real system dy-namics. We can thus rewrite the constraints showing explicitmodel dependence as, A c ( x, f, g ) u ≤ b c ( x, f, g ) . (74)If a controller naively enforces these constraints using thenominal model available to the controller, the controller willenforce the constraint A c ( x, ˜ f , ˜ g ) u ≤ b c ( x, ˜ f , ˜ g ) . (75)Clearly, the constraint in (75) is different from the desiredconstraint we want to enforce on the real model (74). Forinstance, to enforce a contact force constraint, if the controllercomputes and enforces the contact force constraint using thenominal model, there is absolutely no guarantee that the actualcontact force on the real system satisfies the constraint. Remark 9:
As noted earlier, certain constraints do not de-pend on the model at all. In such cases, model uncertainty doesnot affect the constraint. One example of such a constraint isa pure input constraint, such as u ( x ) ≤ u max . Expressing thisconstraint in the form of (74) results in A c = I, b c = u max ,which clearly is not dependent on the model.To robustify the “constraints”, we can use a similar methodas we did for the control barrier functions. We start byreformulating the constraints (74) by using Virtual Input-Output Linearization (VIOL) for to obtain, A c,i ( x, f, g ) u − b c,i ( x, f, g ) =: µ c,i , (76)and then enforce µ c,i ≤ , (77) where the index i indicates the i th constraints in (74).With model uncertainty, we now have, A c,i ( x, f, g ) u − b c,i ( x, f, g ) = µ c,i + ∆ c,i + ∆ c,i µ c,i =: ¯ µ c,i ( µ c,i , ∆ c,i , ∆ c,i ) , (78)with ∆ c,i , ∆ c,i ∈ R and the virtual input µ c,i is now definedbased on the nominal model ˜ f , ˜ g : µ c,i = A c,i ( x, ˜ f , ˜ g ) u − b c,i ( x, ˜ f , ˜ g ) . (79)The robust constraint now becomes ¯ µ c,i ( µ c,i , ∆ c,i , ∆ c,i ) ≤ . (80)Once again, we make the assumption on bounded uncer-tainty, | ∆ c,i | ≤ ∆ c,i ,max , | ∆ c,i | ≤ ∆ c,i ,max , (81)such that the robust constraint condition becomes, max | ∆ c,i |≤ ∆ c,i ,max | ∆ c,i |≤ ∆ c,i ,max ¯ µ c,i ( µ c,i , ∆ c,i , ∆ c,i ) ≤ . (82)Similar to robust CLF and robust CBF, the above robustconstraints condition is affine in µ c,i and can be expressed as max | ∆ c,i |≤ ∆ c,i ,max | ∆ c,i |≤ ∆ c,i ,max Ψ c,i + Ψ c,i µ c,i ≤ (83)where “ ( c, i ) ” refers to the i th constraint, and Ψ c,i ( x, ∆ c,i ) := ∆ c,i , Ψ c,i ( x, ∆ c,i ) := 1 + ∆ c,i , (84)where the above arises due to the form of the constraints in(78). The same procedure as what was done for the robustCBF in the previous section can be carried out for the robustconstraints as well to show how the max inequality getsconverted to a set of linear inequalities and thus the min-maxproblem then becomes a quadratic program even for the robustconstraints as well. E. Robust CBF-CLF-QP with Robust Constraints
We finally can unify the robust CLF for stability, robust CBFfor safety enforcement, and the robust constraints under modeluncertainty to obtain the following unified robust controller. Robust CBF-CLF-QP with Robust Constraints : u ∗ ( x ) = (85) argmin u,µ,µ v ,µ b ,µ c ,δ µ T µ + pδ s.t. max | ∆ v |≤ ∆ v ,max | ∆ v |≤ ∆ v ,max ˙ V ε ( η, ∆ v , ∆ v , µ v ) + c ε V ε ≤ δ,L ˜¯ g V ε µ = µ v , ( Robust CLF ) max | ∆ b |≤ ∆ b ,max | ∆ b |≤ ∆ b ,max ˙ B ( x, ∆ b , ∆ b , µ b ) − γB ( x ) ≤ ,L ˜ f B ( x ) + L ˜ g B ( x ) u = µ b , ( Robust CBF ) max | ∆ c |≤ ∆ c ,max | ∆ c |≤ ∆ c ,max ¯ µ c ( µ c , ∆ c , ∆ c ) ≤ ,A c ( x ) µ − b c ( x ) = µ c , ( Robust Constraints ) u = ˜ u ff ( x ) + ( L ˜ g L ˜ f y ( x )) − µ. ( IO )Having presented our proposed optimal robust controllerthat can address stability and strictly enforce constraints undermodel uncertainty, we now validate our controller in simula-tions on bipedal robot systems and experiments on spring cartsystem.V. S IMULATION AND E XPERIMENTAL V ALIDATION
A. RABBIT Bipedal Robot
To demonstrate the effectiveness of the proposed robustCBF-CLF-QP controller, we will conduct numerical simula-tions on the model of RABBIT, a planar five-link bipedalrobot. Further description of RABBIT and the associatedmathematical model can be found in [8, 50].We consider model uncertainty in bipedal robotic walkingby adding an unknown heavy load to the torso of the RABBITrobot to validate the performance of our proposed robustcontrollers. We will also require enforcement of contact forceconstraints (state constraints) and foot-step location constraints(safety constraints) in the presence of the model uncertainty.In the following simulations, we run an offline optimizationprocess to generate a walking gait for the nominal system.This results in a set of outputs (virtual constraints) that needto be regulated to zero by the controller. Although the offlineoptimization is on the nominal system, as we will see, therobust controller is able to guarantee enforcement of thevirtual constraints on the true model while subject to inputconstraints, contact force constraints, safety constraints, andmodel uncertainty.
1) Dynamic Walking of Bipedal Robot while CarryingUnknown Load, subject to Contact Force Constraints:
Theproblem of contact force constraints is very important forrobotic walking, as any violation of contact constraints wouldcause the robot to slip and fall. We design our nominal walkinggait so as to satisfy contact constraints. However, we cannotguarantee this constraint once there is a perturbation from thenominal walking gait. Although the feedback controller (for example CLF-QP), will drive any error back to the periodicgait, however there is no way to enforce the contact forceconstraint. In our simulation, in addition to model uncertainty,we will introduce a small perturbation in initial configurationof the robot , resulting in an initial CLF V = η T P η = 5 . (see Fig.(1)). We will compare three different controllers, I: CLF-QPII: CLF-QP with Constraints (Contact Force Constraints)III: Robust CLF-QP with Robust Constraints (86)We also enforce input saturation constraints for all threecontroller in (86). However, since this constraint doesn’tdepend on the system model, a robust version for the constraintis not necessary.Three simulation cases with different loads carried on thetorso of the robot are conducted:
Case 1: m load = 0 [kg]Case 2: m load = 5 [kg] (16%)Case 3: m load = 15 [kg] (47%) (87)We consider contact force constraints as follows. Let F hst and F vst be the horizontal and vertical contact force betweenthe stance foot and the ground, in order to avoid slippingduring walking, we will have to guarantee: F vst ( x ) ≥ δ N > (88) (cid:12)(cid:12)(cid:12)(cid:12) F hst ( x ) F vst ( x ) (cid:12)(cid:12)(cid:12)(cid:12) ≤ k f (89)where δ N is a positive threshold for the vertical contact force,and k f is the friction coefficient. In our simulation, we picked δ N = 0 . m robot and k f = 0 . , with m robot = 32[ kg ] beingthe weight of the robot.As we can see from Fig.1, although we just generate asmall initial perturbation, the controller I (CLF-QP) withoutconsidering contact force constraints still violated the frictionconstraint with | F/N | max (cid:39) . , while the controller II (CLF-QP with constraints) can handle this case well. However, witha small model uncertainty (adding m load = 5[ kg ] to the torso),the controller B fails with the friction coefficient | F/N | max (cid:39) . . Interestingly, in this case the robust CLF-QP with robustcontact force constraints controller not only guarantees avery good friction constraints with | F/N | max (cid:39) . , butalso has better tracking performance. With m load = 15[ kg ] ,while the two controllers I (CLF-QP) and II (CLF-QP withconstraints) become unstable and fail in the first walking step,the controller III (Robust CLF-QP with robust constraints)still works well with | k | max (cid:39) . . Especially, we can noticefrom the figures of (cid:107) u (cid:107) that the proposed robust CLF-QP withrobust constraints has a much better performance in both cases,its range of control inputs is nearly the same with those of therest two controllers. In summary, we can conclude that theproposed robust QP offers a novel method that can increasethe robustness of both stability and constraints while usingthe same range of control inputs with prior controllers. Theseproperties will be further strengthened in the next interesting Time (s) C L F Time (s) || u || ( N m ) Time (s) F s t v ( N ) Time (s) | F s t h / F s t v | Case 1: m load = 0 [ kg ] Time (s) C L F Time (s) || u || ( N m ) Time (s) F s t v ( N ) Time (s) | F s t h / F s t v | Case 2: m load = 5 [ kg ] Time (s) C L F Time (s) || u || ( N m ) Time (s) F s t v ( N ) Time (s) | F s t h / F s t v | Case 3: m load = 15 [ kg ] Time (s)
CLF
CLF-QP CLF-QP with Constraints Robust CLF-QP with Robust Constraints
Fig. 1: Bipedal robot with carrying unknown load and contact force constraints. Even in the nominal case of NO uncertainty, theCLF-QP controller fails due to lacking of contact force constraints. The CLF-QP with Constraints (Contact Force Constraints)works well with perfect model but fails with only 5 kg of load. The Robust CLF-QP with Robust Constraints maintains bothgood tracking performance and contact force constraints under up to 15 kg of load (47% of the robot weight). The other twocontrollers are unstable in this case.application for bipedal robotic walking with safety-criticalconstraints.
2) Dynamic Walking of Bipedal Robot while Carrying Un-known Load, subject to Contact Force Constraints and Foot-Step Location Constraints:
For validating our proposed con-troller, we will also test with the problem of footstep locationconstraints when the robot carries an unknown load on thetorso. The control methodology for this problem with perfectmodel can be found in [37]. We will run 100 simulations.For each simulation, the unknown load was chosen randomlybetween 5-15 kg, the desired footstep locations for 10 stepswere chosen randomly between 0.35-0.55 m (the nominalwalking gait has a step length of 0.45 m). Because the CLF-QPcannot handle footstep location constraints, the four followingcontrollers will be compared:
I :CBF-CLF-QP (Foot-Step Placement)II: CBF-CLF-QP with Constraints (Friction Constraints)III: Robust CBF-CLF-QPIV: Robust CBF-CLF-QP with Robust Constraints (90)As you can see from Fig.2, the performance of our proposedrobust CBF-CLF-QP dominates that of the CBF-CLF-QP P e r c en t age o f S u c e ss f u l T e s t s Fig. 2: Dynamic Walking of Bipedal Robot while CarryingUnknown Load, subject to Contact Force Constraints andFoot-Step Location Constraints. 100 random simulations weretested. For each simulation, the unknown load was chooserandomly between 5-15 kg, the desired footstep locationsfor 10 steps were choose randomly between 0.35-0.55 m.The same set of random parameters was tested on the fourcontrollers, where the four controller was specified in (90).(97% of success in comparision with 2%). This result not onlystrengthens the effectiveness of the proposed controller, but italso emphasizes the importance of considering robust controlfor safety constraints because a small model uncertainty cancause violations of such constraints and thereby no longer Fig. 3: Dynamic bipdal walking while carrying unknown load,subject to torque saturation constraints (input constraints), con-tact force constraints (state constraints), and foot-step locationconstraints (safety constraints). Simulation of the Robust CBF-CLF-QP with Robust Constraints controller for walking over10 discrete foot holds is shown, subject to model uncertaintyof 15 Kg (47 %). A video of the simulation is available athttp://youtu.be/tT0xE1XlyDI.
Time (s) h ( x ) h ( x ) Fig. 4: Dynamic walking of bipedal robot while carrying un-known load of 15 Kg (47 %). The CBF constraints, h ( x ) ≥ and h ( x ) ≥ defined in [37], guarantee precise foot-steplocations. The figure depicts data for 10 steps of walking. Ascan be clearly seen, the constraints are strictly enforced despitethe large model uncertainty. Time (s) N ( N ) (a) Vertical Contact Force: N ( x ) > δ N , ( δ N = 0 . mg ) . Time (s) | F / N | (b) Friction Constraint: | F/N | ≤ k f , ( k f = 0 . ) Fig. 5: Dynamic walking of bipedal robot while carryingunknown load of 15 Kg (47 %). (a) Vertical contact forceconstraint and (b) friction constraint are shown for 10 stepswalking. As is evident, both constraints are strictly enforceddespite the large model uncertainty.
Motor EncoderCart (a) Real system
LabVIEW Software(Controller) FPGA boardEncoder MotorRack and PinionSpring-Cart SystemECP BoxEncoder CountsD/A Counts Volts Amps NmNmmEncoder Counts (b) System Diagram
Fig. 6: Schematic of experiment setup for the spring-cartsystem.guaranteeing safety.Figures 3, 4, 5, illustrate one of the runs where the max-imum load of 15 Kg (47% of robot mass) was considered.Stick figure plot, CBF constraints, vertical contact force, andfriction constraint plots are shown. Note that, the simulationswere artificially limited to 10 steps, to enable fast executionof 100 runs for each controller. Simulations for larger numberof steps were also successful as well, but are not presentedhere due to space constraints.
B. Experimental Results on Spring-Cart System
Having presented numerical results of our proposed controlmethod for bipedal robots, we next present experimentalvalidation for the method on a rectilinear spring-cart system,as shown in Fig. 7. It must be noted that the experiments on arectilinear spring-cart system are connected to the simulationswith a bipedal robot. Since our simulations consider nonlinearsystems with IO linearization controllers and safety-constraintswith relative-degree two, our preliminary experiments are thuswith linear systems with relative-degree two safety-constraints.Furthermore, with this experimental result, we will perturbthe system with different types of model uncertainties (seeFig. 7). It therefore can represent model uncertainty in the IOlinearized system discussed in Section IV. Future work willconsider experiments with bipedal systems.For this experiment, our control problem is to track desiredset-point ( x → x d = 2 [ cm ] ) by using a CLF, and guaranteestate-dependent constraint ( x ≤ x max = 1 [ cm ] ) by using aCBF, where x is the position of Cart 1.The experimental setup of the rectilinear spring-cart system(ECP Model 210) is described in Fig 6. Our controller isimplemented using LabVIEW, wherein we call a custom-generated C++ code that implements a fast QP solver. This Time (s) x ( c m ) Case 1: No uncertainty
Time (s) x ( c m ) Case 2: Full-loaded cart 1
Time (s) x ( c m ) Case 3: Case 2 and shaking the table
Time (s) x ( c m ) Case 4: Full-loaded cart 1 + spring +full-loaded cart 2
Time (s) x ( c m ) Case 5: Case 4 and shaking the table
Time (s) x ( c m ) Case 6: Case 4 and perturbing the cart2.
Time (s) x ( c m ) x (CBF) x (Robust CBF) x max x d Fig. 7: Experimental results on the spring-mass system. The goal is to drive the cart to the target location of 2 cm, whileenforcing the safety constraint that the cart does not cross 1 cm. The controller just uses the nominal model as illustratedin Case 1 for all the 6 different cases. Model uncertainty is introduced for Cases 2 in the form of an added unknown mass.Additionally for Case 3, a perturbation is introduced in the form of shaking the table. For Case 4, in addition to the unknownmass, an unkown dynamics is introduced in the form of another cart that is connected through a spring. Additionally for Cases5 and 6, perturbations are introduced in the form of manually shaking the table and shaking the second cart respectively. Inall cases, the proposed robust controller still enforces the strict safety-critical constraint and maintains the cart position under1 cm. A video of the experiment is available at https://youtu.be/g1UewP4R8L4.QP solver code is autogenerated using CVXGEN [31]. Thecontroller runs at 100 Hz on a LabVIEW PXI DAQ andoutputs the control input to the ECP system. In particular,the control input is sent to a FPGA board in the PXI DAQ,which then generates and outputs an analog driving voltage(through a 16-bit Digital-to-Analog Converter) to the currentamplifier in the ECP system. This amplifier generates therequired current to drive the motor which in turn producesa torque. For rectilinear systems this torque is converted toa linear force through a rack and pinion mechanism. Themotion of the moving cart is measured by an encoder andthis information in encoder counts is acquired by the FPGAboard in the PXI DAQ and sent to our controller via the hostLabVIEW software.Fig.7 compares the performance of two controllers CBF-CLF-QP (dotted blue line) and Robust CBF-CLF-QP (red line)under six different cases. Experimental setup for each caseand corresponding result are shown in Fig.7. Note that thetwo controllers were designed based on the nominal modelindicated in case 1 (a single cart) and we generated modeluncertainty from case 2-6 by adding masses, shaking table,adding spring and another cart, etc.From Fig.7, we can observe clearly that while in case 1(without model uncertainty), the two controllers have almostthe same performance, from Cases 2-6, our proposed robustCBF-CLF-QP outperformed the nominal CBF-CLF-QP. To bemore specific, the robust CBF-CLF-QP controller maintainsthe constraint ( x ≤ cm ) ) very well, the nominal CBF-CLF-QP fails in all last 5 cases. C. Discussion
The proposed controller has a few shortcomings. Sincewe are solving for the control input under the worst-case(bounded) model uncertainty assumption, the control could beaggressive. This is a typical drawback of robust controllers.Moreover, as mentioned in Remark 6, we only have local fea-sibility of the QP. In particular, if we increase the bounds of theuncertainty significantly, i.e., large values of ∆ max , ∆ max , theoptimization solves for the control input for the worst case, andcould potentially lead to infeasibility of the QP. Thus, there isa trade-off between robustness and feasibility of the controller.If we choose a small bound on the model uncertainty, it couldlead to poor tracking stability and potential violation of thesafety constraint under mild model uncertainty that exceedsthe bounds. In contrast, if we assume the bound of modeluncertainty being too large the QP could become infeasible.Therefore, in the future, a more formal design of the boundeduncertainty assumption should be explored.VI. C ONCLUSION
We have presented a novel Optimal Robust Control tech-nique that uses control Lyapunov and barrier functions to suc-cessfully handle significantly high model uncertainty for bothstability, input-based constraints, state-dependent constraints,and safety-critical constraints. We validated our proposed con-troller on different problems both numerically and experimen-tally, which show the same property: under model uncertainty,our Robust Control based QP, has much better tracking per-formance and guarantees desired constraints while other typesof QP controllers using Lyapunov and barrier functions not only have large tracking errors but also violate the constraintswith even a small model uncertainty. We show numericalvalidations on dynamically walking of bipedal robots whilesubject to torque saturation and contact force constraints in thepresence of model uncertainty, and dynamically walking withprecise foot placements over a terrain of stepping stones whilesubject to model uncertainty. We also experimentally validateour controller on a spring-cart system subject to significantmodel uncertainty and perturbations. Future directions involveexperimental validations on bipedal robots and other dynamicrobotic systems. A PPENDIX AP ROOF OF T HEOREM ε > be fixed and select a Lipschitz continuousfeedback u ε of the relaxed CLF-QP controller (23). From [2,(56)], we have T εI ( η, z ) is continuous (since it is Lipschitz)and therefore there exists δ > and ∆ T > such that for all ( η, z ) ∈ B δ (0 , ∩ S , T ∗ − ∆ T ≤ T εI ( η, z ) ≤ T ∗ + ∆ T, (91)where T ∗ is the period of the orbit O Z .In order to make use of the proof of the exponential stabilityfor the standard RES-CLF controller in [2], we will presentthe condition for bounding the system states η ( t ) at the time-to-impact T εI ( η, z ) in the following lemma. Lemma 1:
Let O Z be an exponentially stable periodic orbitof the hybrid zero dynamics H | Z (41) transverse to S ∩ Z andthe continuous dynamics of H (8) controlled by a CLF-QPwith relaxed inequality (23) . Then for each ∆ T > and ε > ,there exists an ¯ w ε ≥ such that, if the solution u ε ( η, z ) of theCLF-QP with relaxed inequality (23) satisfies w ε ( T εI ( η, z )) ≤ ¯ w ε , then (cid:107) η ( t ) (cid:107) (cid:12)(cid:12)(cid:12)(cid:12) t = T εI ( η,z ) ≤ (cid:114) c c e − ( T ∗ − ∆ T ) c (cid:107) η (0) (cid:107) . (92) Proof:
From (39) and because w ε ( T εI ( η, z )) ≤ ¯ w ε , itimplies that (cid:107) η ( t ) (cid:107) (cid:12)(cid:12)(cid:12)(cid:12) t = T εI ( η,z ) ≤ (cid:114) c c ε e − c ε T εI ( η,z )+ w ε ( T εI ( η,z )) (cid:107) η (0) (cid:107)≤ (cid:114) c c ε e − c ε T εI ( η,z )+ ¯ w ε (cid:107) η (0) (cid:107) (93)Then, from (91), we have, (cid:107) η ( t ) (cid:107) (cid:12)(cid:12)(cid:12)(cid:12) t = T εI ( η,z ) ≤ (cid:114) c c ε e − c ε ( T ∗ − ∆ T )+ ¯ w ε (cid:107) η (0) (cid:107) . (94) Furthermore, because e − α ≤ e − /α, ∀ α ≥ , we have: ε e − c ε ( T ∗ − ∆ T ) ≤ e − ( T ∗ − ∆ T ) c . (95)Then it implies that there exists a ¯ c ≤ c such that: ε e − c ε ( T ∗ − ∆ T ) ≤ ε e − ¯ c ε ( T ∗ − ∆ T ) ≤ e − ( T ∗ − ∆ T ) c . (96)The above inequality follows by the fact that ε e − c ε ( T ∗ − ∆ T ) is a monotonically decreasing function of c , and that for anyreal numbers a ≤ b , there always exists a number c ∈ [ a : b ] such that a ≤ c ≤ b . To be more specific, let ¯ c a be the solutionof: ε e − ¯ ca ε ( T ∗ − ∆ T ) = 2 e − ( T ∗ − ∆ T ) c , (97)then any ¯ c ∈ [¯ c a : c ] will satisfy (96).We can then define ¯ w ε := c − ¯ c ε ( T ∗ − ∆ T ) ≥ , (98) ⇒ − c ε ( T ∗ − ∆ T ) + 12 ¯ w ε = − ¯ c ε ( T ∗ − ∆ T ) . (99)Next, plugging (99) into (94), we have, (cid:107) η ( t ) (cid:107) (cid:12)(cid:12)(cid:12)(cid:12) t = T εI ( η,z ) ≤ (cid:114) c c ε e − ¯ c ε ( T ∗ − ∆ T ) (cid:107) η (0) (cid:107) , (100) V ε ( η ( t )) (cid:12)(cid:12) t = T εI ( η,z ) ≤ e − ¯ c ε ( T ∗ − ∆ T ) V ε ( η (0)) . (101)We now complete the proof of Lemma 1 by substituting (96)into (100) to obtain (92). The solution of ¯ w ε can be found in(98).Using Lemma 1, we then can follow the same protocol ofthe proof in [2, Theorem 2] until equation [2, (64)].We then define β ( ε ) = c ε L X e − c ε ( T ∗ − ∆ T ) and ¯ β ( ε ) = c ε L X e − ¯ c ε ( T ∗ − ∆ T ) where L ∆ X , defined after [2, (59)],is the Lipschitz constant for ∆ X . Because β (0 + ) :=lim ε (cid:38) β ( ε ) = 0 , then there exists an ε such that β ( ε ) < c ∀ < ε < ε. (102)and for each ε , if we define ¯ c b be the solution of c ε L X e − ¯ cb ε ( T ∗ − ∆ T ) = c , (103)then for ¯ c ∈ (¯ c b : c ] , we have, β ( ε ) ≤ ¯ β ( ε ) < c . (104)However, as presented in proof of Lemma 1, ¯ c also needs tosatisfy (96). Therefore, in order to guarantee the satisfactionof both (96) and (104), ¯ c needs to be chosen in the followingset ¯ c ∈ { [¯ c a : c ] ∩ (¯ c b : c ] } , (105)where ¯ c a and ¯ c b are defined in (97) and (103) respectively.The rest of the proof follows from the proof of [2, Theorem2] using ¯ β instead of β . We finish our proof with the valueof ¯ w ε determined via (98), in which the feasible set of ¯ c isdefined in (105). R EFERENCES [1] A. D. Ames, J. Grizzle, and P. Tabuada, “Control barrierfunction based quadratic programs with application toadaptive cruise control,” in
IEEE Conference on Decisionand Control , 2014, pp. 6271–6278.[2] A. D. Ames, K. Galloway, J. W. Grizzle, and K. Sreenath,“Rapidly Exponentially Stabilizing Control LyapunovFunctions and Hybrid Zero Dynamics,”
IEEE Trans.Automatic Control , vol. 59, no. 4, pp. 876–891, 2014.[3] D. Angeli, E. Sontag, and Y. Wang, “A characterizationof integral input-to-state stability,”
IEEE Transactions onAutomatic Control , vol. 45, pp. 1082–1097, 2000.[4] M. Benallegue and J.-P. Laumond, “Metastability forhigh-dimensional walking systems on stochasticallyrough terrain,” in
Robotics: Science and Systems , 2013.[5] K. Byl and R. Tedrake, “Metastable walking machines,”
The International Journal of Robotics Research , vol. 28,no. 8, pp. 1040–1064, August 2009.[6] C. Cai and A. R. Teel, “Characterizations of input-to-state stability for hybrid systems,”
Systems & ControlLetters , vol. 58, pp. 47–53, 2009.[7] R. Cheng, M. J. Khojasteh, A. D. Ames, and J. W. Bur-dick, “Safe multi-agent interaction through robust con-trol barrier functions with learned uncertainties,” arXivpreprint arXiv:2004.05273 , 2020.[8] C. Chevallereau, G. Abba, Y. Aoustin, F. Plestan, E. R.Westervelt, C. Canudas-de-Wit, and J. W. Grizzle, “RAB-BIT: A testbed for advanced control theory,” vol. 23,no. 5, pp. 57–79, October 2003.[9] R. Deits and R. Tedrake, “Footstep planning on uneventerrain with mixed-integer convex optimization.”
Pro-ceedings of the 2014 IEEE/RAS International Conferenceon Humanoid Robots , pp. 279–286, 2014.[10] S. V. Drakunov and V. I. Utkin, “Sliding mode controlin dynamic systems,”
International Journal of Control ,vol. 55, pp. 1029–1037, 1992.[11] C. Edwards and S. Spurgeon,
Sliding Mode Control:Theory And Applications . CRC Press, 1998.[12] F. Fahimi, “Sliding-mode formation control for underac-tuated surface vessels,”
IEEE Transactions on Robotics ,vol. 23, no. 3, pp. 617–622, 2007.[13] D. D. Fan, J. Nguyen, R. Thakker, N. Alatur, A.-a. Agha-mohammadi, and E. A. Theodorou, “Bayesian learning-based adaptive control for safety critical systems,” arXivpreprint arXiv:1910.02325 , 2019.[14] S. Feng, E. Whitman, X. Xinjilefu, and C. G. Atke-son, “Optimization-based full body control for the darparobotics challenge,”
Journal of Field Robotics , vol. 32,no. 2, pp. 293–312, 2015.[15] R. A. Freeman and P. V. Kokotovi´c,
Robust NonlinearControl Design . Birkhäuser, 1996.[16] D. Fridovich-Keil, S. L. Herbert, J. F. Fisac, S. Deglurkar,and C. J. Tomlin, “Planning, fast and slow: A frame-work for adaptive real-time safe trajectory planning,” arxiv:1710.04731 , 2017.[17] K. Galloway, K. Sreenath, A. D. Ames, and J. W. Grizzle,“Torque saturation in bipedal robotic walking through control lyapunov function based quadratic programs,”
IEEE Access , vol. PP, no. 99, p. 1, April 2015.[18] ——, “Torque saturation in bipedal robotic walkingthrough control lyapunov function based quadratic pro-grams,”
IEEE Access , vol. PP, no. 99, p. 1, April 2015.[19] K. A. Hamed and J. W. Grizzle, “Robust event-basedstabilization of periodic orbits for hybrid systems: Appli-cation to an underactuated 3d bipedal robot,” in
AmericanControl Conference , 2013.[20] ——, “Iterative robust stabilization algorithm for peri-odic orbits of hybrid dynamical systems: Applicationto bipedal running,” in
IFAC Analaysis and Design ofHybrid Systems , 2015.[21] J. P. Hespanhaa, D. Liberzon, and A. R. Teel, “Lyapunovconditions for input-to-state stability of impulsive sys-tems,”
Automatica , vol. 44, pp. 2735–2744, 2008.[22] S.-C. Hsu, X. Xu, and A. D. Ames, “Control barrier func-tion based quadratic programs with application to bipedalrobotic walking,” in
American Control Conference , 2015,pp. 4542–4548.[23] K. Karydis, I. Poulakakis, J. Sun, and H. G. Tanner,“Probabilistically valid stochastic extensions of determin-istic models for systems with uncertainty,”
The Interna-tional Journal of Robotics Research , vol. 34, no. 10, pp.1278–1295, 2015.[24] M. Kelly and A. Ruina, “Non-linear robust control forinverted-pendulum 2d walking,”
IEEE International Con-ference on Robotics and Automation , 2015.[25] H. Khalil,
Nonlinear Systems - 3rd Edition , Upper SaddleRiver, NJ, 2002.[26] S. Kousik, S. Vaskov, M. Johnson-Roberson, and R. Va-sudevan, “Safe trajectory synthesis for autonomousdriving in unforeseen environments,” arXiv:1705.00091 ,2017.[27] F. L. Lewis, D. Vrabie, and V. L. Syrmos,
OptimalControl . John Wiley & Sons, 2012.[28] Z. Li and S. S. Ge, “Adaptive robust controls of bipedrobots,”
IET Control Theory and Applications , vol. 7,no. 2, pp. 161–175, 2013.[29] F. Lin and R. D. Brandt, “An optimal control approach torobust control of robot manipulators,”
IEEE Transactionson Robotics and Automation , vol. 14, no. 1, pp. 69–77,February 1998.[30] A. Majumdar and R. Tedrake, “Funnel libraries for real-time robust feedback motion planning,”
Under Review ,2016.[31] J. Mattingley and S. Boyd, “CVXGEN: a code generatorfor embedded convex optimization,”
Optimization andEngineering , vol. 13, no. 1, pp. 1–27, March 2012.[32] A. Mehra, W.-L. Ma, F. Berg, P. Tabuada, J. W. Grizzle,and A. D. Ames, “Adaptive cruise control: Experimentalvalidation of advanced controllers on scale-model cars,”in
American Control Conference , 2015, pp. 1411–1418.[33] B. Morris, E. Westervelt, C. Chevallereau, G. Buche, andJ. Grizzle,
Fast Motions Symposium on Biomechanicsand Robotics , ser. Lecture Notes in Control and Informa-tion Sciences. Heidelberg, Germany: Springer-Verlag,2006, ch. Achieving Bipedal Running with RABBIT: Six Steps toward Infinity, pp. 277–297.[34] B. Morris, M. J. Powell, and A. D. Ames, “Sufficient con-ditions for the Lipschitz continuity of QP-based multi-objective control of humanoid robots,”
Proc. 52nd IEEEConf. Decision and Control , pp. 2920–2926, 2013.[35] Q. Nguyen and K. Sreenath, “L1 adaptive controlfor bipedal robots withcontrol lyapunov function basedquadratic programs,” in
American Control Conference ,2015, pp. 862–867.[36] ——, “Optimal robust control for bipedal robots throughcontrol lyapunov function based quadratic programs,” in
Robotics: Science and Systems , 2015.[37] ——, “Safety-critical control for dynamical bipedalwalking with precise footstep placement,” in
The IFACConference on Analysis and Design of Hybrid Systems ,2015, pp. 147–154.[38] ——, “Exponential control barrier functions for enforc-ing high relative-degree safety-critical constraints,” in
American Control Conference , 2016.[39] ——, “Optimal robust control for constrained nonlinearhybrid systems with application to bipedal locomotion,”in
American Control Conference , 2016.[40] S. J. Qin and T. A. Badgwell, “A survey of industrialmodel predictive control technology,”
Control engineer-ing practice , vol. 11, no. 7, pp. 733–764, 2003.[41] J. S. Reed and P. A. Ioannou, “Instability analysis androbust adaptive control of robotic manipulators,”
IEEEtransactions on robotics and automation , vol. 5, no. 3,pp. 381–386, 1989.[42] C. O. Saglam and K. Byl, “Meshing hybrid zero dy-namics for rough terrain walking,” in
IEEE Interna-tional Conference on Robotics and Automation , 2015,pp. 5718–5725.[43] E. D. Sontag, “A ’universal’ construction of Artsteintheorem on nonlinear stabilization,”
Systems & ControlLetters , vol. 13, pp. 117–123, 1989.[44] ——, “On the input-to-state stability property,”
EuropeanJournal of Control , vol. 1, pp. 24–36, 1995.[45] E. D. Sontag and Y. Wang, “On characterizations ofthe input-to-state stability property,”
Systems & ControlLetters , vol. 24, pp. 351–359, 1995.[46] M. W. Spong, “On the robust control of robot manipula-tors,”
IEEE Transactions on Automatic Control , vol. 37,no. 11, pp. 1782–1786, November 1992.[47] C.-Y. Su and T.-P. Leung, “A sliding mode controllerwith bound estimation for robot manipulators,”
IEEETransactions on Robotics and Automation , vol. 9, no. 2,pp. 208–214, 1993.[48] L. Vu, D. Chatterjee, and D. Liberzon, “Input-to-statestability of switched systems and switching adaptivecontrol,”
Automatica , vol. 43, pp. 639–646, 2007.[49] Y. Wang and S. Boyd, “Fast model predictive controlusing online optimization,”
IEEE Transactions on controlsystems technology , vol. 18, no. 2, pp. 267–278, 2010.[50] E. R. Westervelt, G. Buche, and J. W. Grizzle, “Ex-perimental validation of a framework for the design ofcontrollers that induce stable walking in planar bipeds,”vol. 24, no. 6, pp. 559–582, June 2004. [51] G. Wu and K. Sreenath, “Safety-critical and constrainedgeometric control synthesis using control lyapunov andcontrol barrier functions for systems evolving on mani-folds,” in
American Control Conference , 2015, pp. 2038–2044.[52] X. Xu, P. Tabuada, J. W. Grizzle, and A. D. Ames,“Robustness of control barrier functions for safety criticalcontrol,” in
The IFAC Conference on Analysis and Designof Hybrid Systems , 2015, pp. 54–61.[53] K. Zhou, J. Doyle, and K. Glover,
Robust and OptimalControl . Prentice Hall PTR, 1996.
PLACEPHOTOHERE
Quan Nguyen is an Assistant Professor ofAerospace and Mechanical Engineering at the Uni-versity of Southern California. Prior to joining USC,he was a Postdoctoral Associate in the BiomimeticRobotics Lab at the Massachusetts Institute ofTechnology (MIT). He received his Ph.D. fromCarnegie Mellon University (CMU) in 2017 with theBest Dissertation Award. His research interests spandifferent control and optimization approaches forhighly dynamic robotics including nonlinear control,trajectory optimization, real-time optimization-basedcontrol, robust and adaptive control.PLACEPHOTOHERE