Ulrik Pagh Schultz

A distributed and morphology-independent strategy for adaptive locomotion in self-reconfigurable modular robots

In this paper, we present a distributed reinforcement learning strategy for morphology-independent life-long gait learning for modular robots. All modules run identical controllers that locally and independently optimize their action selection based on the robots velocity as a global, shared reward signal. We evaluate the strategy experimentally mainly on simulated, but also on physical, modular robots. We find that the strategy: (i) for six of seven configurations (3-12 modules) converge in 96% of the trials to the best known action-based gaits within 15 min, on average, (ii) can be transferred to physical robots with a comparable performance, (iii) can be applied to learn simple gait control tables for both M-TRAN and ATRON robots, (iv) enables an 8-module robot to adapt to faults and changes in its morphology, and (v) can learn gaits for up to 60 module robots but a divergence effect becomes substantial from 20-30 modules. These experiments demonstrate the advantages of a distributed learning strategy for modular robots, such as simplicity in implementation, low resource requirements, morphology independence, reconfigurability, and fault tolerance.

A distributed strategy for gait adaptation in modular robots

In this paper we study online gait optimization for modular robots. The learning strategy we apply is distributed, independent on robot morphology, and easy to implement. First we demonstrate how the strategy allows an ATRON robot to adapt to faults and changes in its morphology and we study the strategys scalability. Second we extend the strategy to learn the parameters of gait-tables for ATRON and M-TRAN robots.We conclude that the presented strategy is effective for online learning of gaits for most types of modular robots and that learning can effectively be distributed by having independent processes learning in parallel.

Fractal gene regulatory networks for robust locomotion control of modular robots

Designing controllers for modular robots is difficult due to the distributed and dynamic nature of the robots. In this paper fractal gene regulatory networks are evolved to control modular robots in a distributed way. Experiments with different morphologies of modular robot are performed and the results show good performance compared to previous results achieved using learning methods. Furthermore, some experiments are performed to investigate evolvability of the achieved solutions in the case of module failure and it is shown that the system is capable of come up with new effective solutions.

Robust and reversible execution of self-reconfiguration sequences???

Modular, self-reconfigurable robots are robotic systems that can change their own shape by autonomously rearranging the physical modules from which they are built. In this work, we are interested in how to distributedly execute a specified self-reconfiguration sequence. The sequence is specified using a simple and centralized scripting language, which either could be the outcome of a planner or be hand-coded. The distributed controller generated from this language allows for parallel self-reconfiguration steps and is highly robust to communication errors and loss of local state due to software failures. Furthermore, the self-reconfiguration sequence can automatically be reversed, if desired. We verify our approach and demonstrate its robustness in experiments using physical and the simulated ATRON modules, as well as simulated M-TRAN modules. Overall, the contribution of this work is the combination of the tractability of a centralized scripting language with the robustness and parallelism of distributed controllers in modular robots.

Spatial Computing: Distributed Systems That Take Advantage of Our Geometric World

The modeling and control of systems composed of many computational devices is a perennial problem. This problem is growing more acute as the number and density of computing devices continues to rise rapidly. At the macro-scale, the number of computers per person continues to shoot upwards—from traditional PCs and cell-phones to appliances and consumer products to sensor networks and unmanned vehicles— and better and more pervasive networking binds them into larger aggregates. At the micro-scale, the number of devices that can be packed onto a chip continues to climb, and emerging platforms in areas such as nanotechnology and synthetic biology offer the potential to cheaply create systems of billions or trillions of semi-reliable devices. This trend even extends into the natural world, as we learn more about the complex computations carried out by aggregates of living organisms, such as the cells comprising an organism or a biofilm. In recent years, spatial computing has emerged as a promising approach to the modeling and control of these sorts of aggregate systems. The basic insight of spatial computing is simple: when the density of computing devices is high, there is a close relationship between the structure of the network of devices and the geometry of the space through which they are distributed. Put more formally: a spatial computer is any aggregate of devices in which the difficulty of moving information between any two devices is strongly dependent on the distance between them, and the functional goals of the system are generally defined in terms of the system’s spatial structure. This insight gives power in two ways: first, geometric models often enable elegant solutions to problems of robustness, adaptability, scalability, and coordination. Second, the common spatial model unites problems across a wide variety of domains, allowing results to be transferred between them.

Model-based kinematics generation for modular mechatronic toolkits

Modular robots are mechatronic devices that enable the construction of highly versatile and flexible robotic systems whose mechanical structure can be dynamically modified. The key feature that enables this dynamic modification is the capability of the individual modules to connect to each other in multiple ways and thus generate a number of different mechanical systems, in contrast with the monolithics fixed structure of conventional robots. The mechatronic flexibility, however, complicates the development of models and programming abstractions for modular robots, since manually describing and enumerating the full set of possible interconnections is tedious and error-prone for real-world robots. In order to allow for a general formulation of spatial abstractions for modular robots and to ensure correct and streamlined generation of code dependent on mechanical properties, we have developed the Modular Mechatronics Modelling Language (M3L). M3L is a domain-specific language, which can model the kinematic structure of individual robot modules and declaratively describe their possible interconnections rather than requiring the user to enumerate them in their entirety. From this description, the M3L compiler generates the code that is needed to simulate the resulting robots within Webots, widely used commercial robot simulator, and the software component needed for spatial structure computations by a virtual machine-based runtime system, which we have developed and used for programming physical modular robots

Generalized programming of modular robots through kinematic configurations

The distinctive feature of modular robots consists in their reconfigurable mechanical structure, as they are assembled on-demand from basic mechatronic units. This implies that kinematic models of the robots need to be computed on a case-by-case basis for each specific assembly, which is a manual and hence time-consuming and error-prone procedure. We propose to automate this process by automatically computing such kinematic models starting from simple descriptions of the modules and their assemblies. This automated computation is supported by our toolchain for programming arbitrary modular robots in arbitrary configurations, presented in this paper. We contribute two novel results through this approach. First, a high-level programming language that provides kinematic abstractions for arbitrary modular robots, in contrast to the robot-specific solutions currently available. Second, a programming abstraction to subsume multiple kinematically equivalent robot assemblies into a so-called kinematic configuration, hence eliminating the need to explicitly enumerate and program each of them. These contributions advance current techniques for modular robot programming by demonstrating a tool that a) targets multiple mechanical platforms, offering the first general solution for modular robot programming, and b) raises the abstraction level by allowing users to reason and program in terms of standardized kinematic models that are automatically mapped to physical robot configurations by the toolchain.

Towards a Domain-Specific Language for Reversible Assembly Sequences

Programming industrial robots for small-sized batch production of assembly operations is challenging due to the difficulty of precisely specifying general yet robust assembly operations. We observe that as the complexity of assembly increases, so does the likelihood of errors. We propose that certain classes of errors during assembly operations can be addressed using reverse execution, allowing the robot to temporarily back out of an erroneous situation, after which the assembly operation can be automatically retried. Moreover, reversibility can be used to automatically derive a disassembly sequence from a given assembly sequence, or vice versa.

Towards Rule-Based Dynamic Safety Monitoring for Mobile Robots

Safety is a key challenge in robotics, in particular for mobile robots operating in an open and unpredictable environment. To address the safety challenge, various software-based approaches have been proposed, but none of them provide a clearly specified and isolated safety layer. In this paper, we propose that safety-critical concerns regarding the robot software be explicitly declared separately from the main program, in terms of externally observable properties of the software. Concretely, we use a Domain-Specific Language (DSL) to declaratively specify a set of safety-related rules that the software must obey, as well as corresponding corrective actions that trigger when rules are violated. Our prototype DSL is integrated with ROS, is shown to be capable of specifying safety-related constraints, and is experimentally demonstrated to enforce safety behaviour in existing robot software. We believe our approach could be extended to other fields to similarly simplify safety certification.

On the Use of Safety Certification Practices in Autonomous Field Robot Software Development: A Systematic Mapping Study

Robotics has recently seen an increasing development, and the areas addressed within robotics has extended into domains we consider safety-critical, fostering the development of standards that facilitate the development of safe robots. Safety standards describe concepts to maintain desired reactions or performance in malfunctioning systems, and influence industry regarding software development and project management. However, academia seemingly did not reach the same degree of utilisation of standards. This paper presents the findings from a systematic mapping study in which we study the state-of-the-art in developing software for safety-critical software for autonomous field robots. The purpose of the study is to identify practices used for the development of autonomous field robots and how these practices relate to available safety standards. Our findings from reviewing 49 papers show that standards, if at all, are barely used. The majority of the papers propose various solutions to achieve safety, and about half of the papers refer to non-standardised approaches that mainly address the methodical rather than the development level. The present study thus shows an emerging field still on the quest for suitable approaches to develop safety-critical software, awaiting appropriate standards for this support.