Motion capture and reinforcement learning of dynamically stable humanoid movement primitives

Direct transfer of human motion trajectories to humanoid robots does not result in dynamically stable robot movements due to the differences in human and humanoid robot kinematics and dynamics. We developed a system that converts human movements captured by a low-cost RGB-D camera into dynamically stable humanoid movements. The transfer of human movements occurs in real-time. As need arises, the developed system can smoothly transition between unconstrained movement imitation and imitation with balance control, where movement reproduction occurs in the null space of the balance controller. The developed balance controller is based on an approximate model of the robot dynamics, which is sufficient to stabilize the robot during on-line imitation. However, the resulting movements cannot be guaranteed to be optimal because the model of the robot dynamics is not exact. The initially acquired movement is therefore subsequently improved by model-free reinforcement learning, both with respect to the accuracy of reproduction and balance control. We present experimental results in simulation and on a real humanoid robot.

Imitation Learning of Motion Coordination in Robots

Elena Gribovskaya

The ease with which humans coordinate all their limbs is fascinating. Such a simplicity is the result of a complex process of motor coordination, i.e. the ability to resolve the biomechanical redundancy in an efficient and repeatable manner. Coordination enables a wide variety of everyday human activities from filling in a glass with water to pair figure skating. Therefore, it is highly desirable to endow robots with similar skills. Despite the apparent diversity of coordinated motions, all of them share a crucial similarity: these motions are dictated by underlying constraints. The constraints shape the formation of the coordination patterns between the different degrees of freedom. Coordination constraints may take a spatio-temporal form; for instance, during bimanual object reaching or while catching a ball on the fly. They also may relate to the dynamics of the task; for instance, when one applies a specific force profile to carry a load. In this thesis, we develop a framework for teaching coordination skills to robots. Coordination may take different forms, here, we focus on teaching a robot intra-limb and bimanual coordination, as well as coordination with a human during physical collaborative tasks. We use tools from well-established domains of Bayesian semiparametric learning (Gaussian Mixture Models and Regression, Hidden Markov Models), nonlinear dynamics, and adaptive control. We take a biologically inspired approach to robot control. Specifically, we adopt an imitation learning perspective to skill transfer, that offers a seamless and intuitive way of capturing the constraints contained in natural human movements. As the robot is taught from motion data provided by a human teacher, we exploit evidence from human motor control of the temporal evolution of human motions that may be described by dynamical systems. Throughout this thesis, we demonstrate that the dynamical system view on movement formation facilitates coordination control in robots. We explain how our framework for teaching coordination to a robot is built up, starting from intra-limb coordination and control, moving to bimanual coordination, and finally to physical interaction with a human. The dissertation opens with the discussion of learning discrete task-level coordination patterns, such as spatio-temporal constraints emerging between the two arms in bimanual manipulation tasks. The encoding of bimanual constraints occurs at the task level and proceeds through a discretization of the task as sequences of bimanual constraints. Once the constraints are learned, the robot utilizes them to couple the two dynamical systems that generate kinematic trajectories for the hands. Explicit coupling of the dynamical systems ensures accurate reproduction of the learned constraints, and proves to be crucial for successful accomplishment of the task. In the second part of this thesis, we consider learning one-arm control policies. We present an approach to extracting non-linear autonomous dynamical systems from kinematic data of arbitrary point-to-point motions. The proposed method aims to tackle the fundamental questions of learning robot coordination: (i) how to infer a motion representation that captures a multivariate coordination pattern between degrees of freedom and that generalizes this pattern to unseen contexts; (ii) whether the policy learned directly from demonstrations can provide robustness against spatial and temporal perturbations. Finally, we demonstrate that the developed dynamical system approach to coordination may go beyond kinematic motion learning. We consider physical interactions between a robot and a human in situations where they jointly perform manipulation tasks; in particular, the problem of collaborative carrying and positioning of a load. We extend the approach proposed in the second part of this thesis to incorporate haptic information into the learning process. As a result, the robot adapts its kinematic motion plan according to human intentions expressed through the haptic signals. Even after the robot has learned the task model, the human still remains a complex contact environment. To ensure robustness of the robot behavior in the face of the variability inherent to human movements, we wrap the learned task model in an adaptive impedance controller with automatic gain tuning. The techniques, developed in this thesis, have been applied to enable learning of unimanual and bimanual manipulation tasks on the robotics platforms HOAP-3, KATANA, and i-Cub, as well as to endow a pair of simulated robots with the ability to perform a manipulation task in the physical collaboration.

EPFL2012

Rich and Robust Bio-Inspired Locomotion Control for Humanoid Robots

Nicolas Benoît Dominique Van der Noot

Bipedal locomotion is a challenging task in the sense that it requires to maintain dynamic balance while steering the gait in potentially complex environments. Yet, humans usually manage to move without any apparent difficulty, even on rough terrains. This requires a complex control scheme which is far from being understood. In this thesis, we take inspiration from the impressive human walking capabilities to design neuromuscular controllers for humanoid robots. More precisely, we control the robot motors to reproduce the action of virtual muscles commanded by stimulations (i.e. neural signals), similarly to what is done during human locomotion. Because the human neural circuitry commanding these muscles is not completely known, we make hypotheses about this control scheme to simplify it and progressively refine the corresponding rules. This thesis thus aims at developing new walking algorithms for humanoid robots in order to obtain fast, human-like and energetically efficient gaits. In particular, gait robustness and richness are two key aspects of this work. In other words, the gaits developed in the thesis can be steered by an external operator, while being resistant to external perturbations. This is mainly tested during blind walking experiments on COMAN, a 95 cm tall humanoid robot. Yet, the proposed controllers can be adapted to other humanoid robots. In the beginning of this thesis, we adapt and port an existing reflex-based neuromuscular model to the real COMAN platform. When tested in a 2D simulation environment, this model was capable of reproducing stable human-like locomotion. By porting it to real hardware, we show that these neuromuscular controllers are viable solutions to develop new controllers for robotics locomotion. Starting from this reflex-based model, we progressively iterate and transform the stimulation rules to add new features. In particular, gait modulation is obtained with the inclusion of a central pattern generator (CPG), a neural circuit capable of producing rhythmic patterns of neural activity without receiving rhythmic inputs. Using this CPG, the 2D walker controllers are incremented to generate gaits across a range of forward speeds close to the normal human one. By using a similar control method, we also obtain 2D running gaits whose speed can be controlled by a human operator. The walking controllers are later extended to 3D scenarios (i.e. no motion constraint) with the capability to adapt both the forward speed and the heading direction (including steering curvature). In parallel, we also develop a method to automatically learn stimulation networks for a given task and we study how flexible feet affect the gait in terms of robustness and energy efficiency. In sum, we develop neuromuscular controllers generating human-like gaits with steering capabilities. These controllers recruit three main components: (i) virtual muscles generating torque references at the joint level, (ii) neural signals commanding these muscles with reflexes and CPG signals, and (iii) higher level commands controlling speed and heading. Interestingly, these developments target humanoid robots locomotion but can also be used to better understand human locomotion. In particular, the recruitment of a CPG during human locomotion is still a matter open to debate. This question can thus benefit from the experiments performed in this thesis.

EPFL2017

Using reinforcement learning for optimizing the reproduction of tasks in robot programming by demonstration

Florent Günter

As robots start pervading human environments, the need for new interfaces that would simplify human-robot interaction has become more pressing. Robot Programming by Demonstration (RbD) develops intuitive ways of programming robots, taking inspiration in strategies used by humans to transmit knowledge to apprentices. The user-friendliness of RbD is meant to allow lay users with no prior knowledge in computer science, electronics or mechanics to train robots to accomplish tasks the same way as they would with a co-worker. When a trainer teaches a task to a robot, he/she shows a particular way of fulfilling the task. For a robot to be able to learn from observing the trainer, it must be able to learn what the task entails (i.e. answer the so-called "What-to-imitate?" question), by inferring the user's intentions. But most importantly, the robot must be able to adapt its own controller to fit at best the demonstration (the so-called "How-to-imitate?" question) despite different setups and embodiments. The latter is the question that interested us in this thesis. It relates to the problem of optimizing the reproduction of the task under environmental constraints. The "How-to-imitate?" question is subdivided into two problems. The first problem, also known as the "correspondence problem", relates to resolving the discrepancy between the human demonstrator and robot's body that prevent the robot from doing an identical reproduction of the task. Even though we helped ourselves by considering solely humanoid platforms, that is platforms that have a joint configuration similar to that of the human, discrepancies in the number of degrees of freedom and range of motion remained. We resolved these by exploiting the redundant information conveyed through the demonstrations by collecting data through different frames of reference. By exploiting these redundancies in an algorithm comparable to the damped least square algorithm, we are able to reproduce a trajectory that minimizes the error between the desired trajectory and the reproduced trajectory across each frame of reference. The second problem consists in reproducing a trajectory in an unknown setup while respecting the task constraints learned during training. When the information learned from the demonstration no longer suffice to generalize the task constraints to a new set-up, the robot must re-learn the task; this time through trial-and-error. Here we considered the combination of trial-and-error learning to complement RbD. By adding a trial-and-error module to the original Imitation Learning algorithm, the robot can find a solution that is more adapted to the context and to its embodiment than the solution found using RbD. Specifically, we compared Reinforcement Learning (RL) – to other classical optimization techniques. We show that the system is advantageous in that: a) learning is more robust to unexpected events that have not been encountered during the demonstrations and b) the robot is able to optimize its own model of the task according to its own embodiment.

EPFL2009