Harish Chaandar Ravichandar

Learning Contracting Nonlinear Dynamics From Human Demonstration for Robot Motion Planning

In this paper, we present an algorithm to learn the dynamics of human arm motion from the data collected from human actions. Learning the motion plans from human demonstrations is essential in making robot programming possible by nonexpert programmers as well as realizing human-robot collaboration. The highly complex human reaching motion is generated by a stable closed-loop dynamical system. To capture the complexity a neural network (NN) is used to represent the dynamics of the human motion states. The trajectories of arm generated by humans for reaching to a place are contracting towards the goal location from various initial conditions with built in obstacle avoidance. To take into consideration the contracting nature of the human motion dynamics the unknown motion model is learned using a NN subject to contraction analysis constraints. To learn the NN parameters an optimization problem is formulated by relaxing the non-convex contraction constraints to Linear matrix inequality (LMI) constraints. Sequential Quadratic Programming (SQP) is used to solve the optimization problem subject to the LMI constraints. For obstacle avoidance a negative gradient of the repulsive potential function is added to the learned contracting NN model. Experiments are conducted on Baxter robot platform to show that the robot can generate reaching paths from the contracting NN dynamics learned from human demonstrated data recorded using Microsoft Kinect sensor. The algorithm is able to adapt to situations for which the demonstrations are not available, e.g., an obstacle placed in the path.Copyright

Human intention inference through interacting multiple model filtering

We present an algorithm to learn human arm motion from demonstrations and infer the goal location (intention) of human reaching actions. To capture the complexity of human arm reaching motion, an artificial neural network (ANN) is used to represent the arm motion dynamics. The trajectories of the arm motion for reaching operation are modeled as stable dynamic systems with contracting behavior towards the goal location. The ANN is trained subjected to contraction analysis constraints. To adapt the motion model learned from a few demonstrations to novel scenarios or multiple objects, we use an interacting multiple model framework. The multiple models are obtained by translating the origin of the contracting system to different known goal locations. The posterior probabilities of the models are calculated through interactive model matched filtering carried out using extended Kalman filters (EKFs). The correct model is chosen according to the posterior probabilities to infer the correct intention. Demonstrations and measurements are recorded using a Microsoft Kinect sensor and experimental results are presented to validate the proposed algorithm.

Human Intention Inference Using Expectation-Maximization Algorithm With Online Model Learning

An algorithm called adaptive-neural-intention estimator (ANIE) is presented to infer the intent of a human operator’s arm movements based on the observations from a 3-D camera sensor (Microsoft Kinect). Intentions are modeled as the goal locations of reaching motions in 3-D space. Human arm’s nonlinear motion dynamics are modeled using an unknown nonlinear function with intentions represented as parameters. The unknown model is learned by using a neural network. Based on the learned model, an approximate expectation-maximization algorithm is developed to infer human intentions. Furthermore, an identifier-based online model learning algorithm is developed to adapt to the variations in the arm motion dynamics, the motion trajectory, the goal locations, and the initial conditions of different human subjects. The results of experiments conducted on data obtained from different users performing a variety of reaching motions are presented. The ANIE algorithm is compared with an unsupervised Gaussian mixture model algorithm and an Euclidean distance-based approach by using Cornell’s CAD-120 data set and data collected in the Robotics and Controls Laboratoy at UConn. The ANIE algorithm is compared with the inverse LQR and ATCRF algorithms using a labeling task carried out on the CAD-120 data set.

Learning periodic motions from human demonstrations using transverse contraction analysis

In this paper, an algorithm called transverse contracting dynamic system primitive (CDSP) to learn the dynamics of periodic motions from demonstrations is presented. Learning motion plans is essential in making robot programming possible by non-expert programmers as well as realizing human-robot collaboration. The complex periodic motion of the human arm is generated by an orbitally stable closed-loop dynamical system. To capture the complexity, a neural network (NN) model is used to represent the dynamics of the human arm motion states. To take into consideration the orbitally stable nature of the human arm motion dynamics, the unknown motion model learning is subjected to transverse contraction analysis constraints. To learn the model parameters, an optimization problem is formulated by relaxing the non-convex transverse contraction constraints using sum of squares (SOS) programming. The NN model is approximated by using a polynomial approximation of the activation function in order to pose the optimization problem as an SOS problem. A negative gradient of a repulsive potential function is added to the learned transverse contracting model to incorporate obstacle avoidance. Experimental results indicate that the model learning via CDSP can generate periodic motions learned from human demonstrated data recorded using Microsoft Kinect sensor. The CDSP algorithm is able to adapt to situations for which the demonstrations are not available, e.g., an obstacle placed in the path. The comparison between the CDSP algorithm and rhythmic dynamic movement primitives (DMPs) shows that the CDSP algorithm performs better in terms of periodic trajectory reproduction accuracy and the time taken by the reproductions to enter the periodic orbit from random initial conditions.

Human intention inference and motion modeling using approximate E-M with online learning

In this paper, we present an algorithm to infer the intent of a human operators arm movements based on the observations from a Microsoft Kinect sensor. Intentions are modeled as goal locations in 3-dimensional (3D) space where the human is intending to reach. Human intention inference is a critical step towards realizing safe human-robot collaboration. This work models the human arms nonlinear motion dynamics using an unknown nonlinear function with intentions modeled as parameters. The unknown model is learned using a neural network (NN). Based on the learned model, an approximate expectation-maximization (E-M) algorithm is developed to infer human intentions. Furthermore, an identifier-based online model learning algorithm is developed to adapt to variations in the arm motion dynamics, trajectory of motion, goal locations, and initial conditions of different human subjects. We show the results of our algorithm using two sets of experiments conducted on data obtained from different users.

Gyro-aided image-based tracking using mutual information optimization and user inputs

In this paper, a template tracking algorithm is presented that uses mutual information (MI) criteria for template matching and gyroscope information to predict rotation between two camera images. The tracking algorithm can also take an user input for template selection and update. The tracking algorithm uses Hu moments that are invariant to 2D rotation, translation and scaling to validate the tracker. Homography is used to represent template warping parameters. The algorithm is aided with gyroscope measurements to estimate the camera motion information which helps to improve the initial guess of the warping condition. Template selection using the users input is based on properties of the target, such as its location in the frame. The user driven strategy makes the tracker capable of tracking different objects of interest and might reduce the computational burden for template localization. The tracking algorithm presented in this paper shows significant improvements over recently developed gyro-aided Kanade-Lucas-Tomasi (KLT) tracker and the MI-only tracker in the case of multi-modal images and rapid camera motion.

Gaze and motion information fusion for human intention inference

An algorithm, named gaze-based multiple model intention estimator (G-MMIE), is presented for early prediction of the goal location (intention) of human reaching actions. The trajectories of the arm motion for reaching tasks are modeled by using an autonomous dynamical system with contracting behavior towards the goal location. To represent the dynamics of human arm reaching motion, a neural network (NN) is used. The parameters of the NN are learned under constraints derived based on contraction analysis. The constraints ensure that the trajectories of the dynamical system converge to a single equilibrium point. In order to use the motion model learned from a few demonstrations in new scenarios with multiple candidate goal locations, an interacting multiple-model (IMM) framework is used. For a given reaching motion, multiple models are obtained by translating the equilibrium point of the contracting system to different known candidate locations. Hence, each model corresponds to the reaching motion that ends at the respective candidate location. Further, since humans tend to look toward the location they are reaching for, prior probabilities of the goal locations are calculated based on the information about the human’s gaze. The posterior probabilities of the models are calculated through interacting model matched filtering. The candidate location with the highest posterior probability is chosen to be the estimate of the true goal location. Detailed quantitative evaluations of the G-MMIE algorithm on two different datasets involving 15 subjects, and comparisons with state-of-the-art intention inference algorithms are presented.

Learning and synchronization of movement primitives for bimanual manipulation tasks

Bimanual manipulation tasks require strong coordination between the two hands performing the task. Trajectories of both arms must be temporally synchronized while satisfying certain spatial constraints while performing the task. In this paper, control laws for the desired trajectory tracking and synchronization of multiple agents modeled using dynamic movement primitives (DMPs) are developed. The control laws are developed using contraction analysis of nonlinear systems. Specific control laws and tracking and synchronization constraints for bimanual tasks are developed. Experimental results suggest that the proposed control laws are robust to spatial perturbations and on-the fly goal location changes.

Expectation Maximization Method to Identify an Electrically Stimulated Musculoskeletal Model

A system identification algorithm for a musculoskeletal system using an approximate expectation maximization (E-M) is presented. Effective control design for neuroprosthesis applications necessitates a well defined muscle model. A dynamic model of the lower leg with a fixed ankle is considered. The unknown parameters of the model are estimated using an approximate E-M algorithm based on knee angle measurements collected from an able-bodied subject during stimulated knee extension. The parameters estimated from the data are compared to reference values obtained by conducting experiments that separate the parameters in the dynamics from one another. The presented results demonstrate the capability of the proposed algorithm to identify the parameters of the dynamic model from knee angle measurements.Copyright