Abstract:Robotic systems are often complex and depend on the integration of a large number of software components. One important component in robotic systems provides the calculation of forward kinematics, which is required by both motion-planning and perception related components. End-to-end learning systems based on deep learning require passing gradients across component boundaries.Typical software implementations of forward kinematics are not differentiable, and thus prevent the construction of gradient-based, end-to-end learning systems. In this paper we present a library compatible with ROS-URDF that computes forward kinematics while simultaneously giving access to the gradients w.r.t. joint configurations and model parameters, allowing gradient-based learning and model identification. Our Python library is based on Tensorflow~2 and is auto-differentiable. It supports calculating a large number of kinematic configurations on the GPU in parallel, yielding a considerable performance improvement compared to sequential CPU-based calculation. https://github.com/lumoe/dlkinematics.git
Abstract:Sufficient exploration is paramount for the success of a reinforcement learning agent. Yet, exploration is rarely assessed in an algorithm-independent way. We compare the behavior of three data-based, offline exploration metrics described in the literature on intuitive simple distributions and highlight problems to be aware of when using them. We propose a fourth metric,uniform relative entropy, and implement it using either a k-nearest-neighbor or a nearest-neighbor-ratio estimator, highlighting that the implementation choices have a profound impact on these measures.
Abstract:Local policy search is performed by most Deep Reinforcement Learning (D-RL) methods, which increases the risk of getting trapped in a local minimum. Furthermore, the availability of a simulation model is not fully exploited in D-RL even in simulation-based training, which potentially decreases efficiency. To better exploit simulation models in policy search, we propose to integrate a kinodynamic planner in the exploration strategy and to learn a control policy in an offline fashion from the generated environment interactions. We call the resulting model-based reinforcement learning method PPS (Planning for Policy Search). We compare PPS with state-of-the-art D-RL methods in typical RL settings including underactuated systems. The comparison shows that PPS, guided by the kinodynamic planner, collects data from a wider region of the state space. This generates training data that helps PPS discover better policies.