The general stochastic optimal control (SOC) problem in robotics scenarios is often too complex to be solved exactly and in near real time. A classical approximate solution is to first compute an optimal (deterministic) trajectory and then solve a local linear-quadratic-gaussian (LQG) perturbation model to handle the system stochasticity. We present a new algorithm for this approach which improves upon previous algorithms like iLQG. We consider a probabilistic model for which the maximum likelihood (ML) trajectory coincides with the optimal trajectory and which, in the LQG case, reproduces the classical SOC solution. The algorithm then utilizes approximate inference methods (similar to expectation propagation) that efficiently generalize to non-LQG systems. We demonstrate the algorithm on a simulated 39-DoF humanoid robot.

No abstract

Abstract-We present a reformulation of the stochastic optimal control problem in terms of KL divergence minimisation, not only providing a unifying perspective of previous approaches in this area, but also demonstrating that the formalism leads to novel practical approaches to the control problem. Specifically, a natural relaxation of the dual formulation gives rise to exact iterative solutions to the finite and infinite horizon stochastic optimal control problem, while direct application of Bayesian inference methods yields instances of risk sensitive control. We furthermore study corresponding formulations in the reinforcement learning setting and present model free algorithms for problems with both discrete and continuous state and action spaces. Evaluation of the proposed methods on the standard Gridworld and Cart-Pole benchmarks verifies the theoretical insights and shows that the proposed methods improve upon current approaches.

We consider the problem of sequential manipulation and tool-use planning in domains that include physical interactions such as hitting and throwing. The approach integrates a Task And Motion Planning formulation with primitives that either impose stable kinematic constraints or differentiable dynamical and impulse exchange constraints at the path optimization level. We demonstrate our approach on a variety of physical puzzles that involve tool use and dynamic interactions. We then compare manipulation sequences generated by our approach to human actions on analogous tasks, suggesting future directions and illuminating current limitations.

Inference in Markov DecisionProcesses has recently received interest as a means to infer goals of an observed action, policy recognition, and also as a tool to compute policies. A particularly interesting aspect of the approach is that any existing inference technique in DBNs now becomes available for answering behavioral questions-including those on continuous, factorial, or hierarchical state representations. Here we present an Expectation Maximization algorithm for computing optimal policies. Unlike previous approaches we can show that this actually optimizes the discounted expected future return for arbitrary reward functions and without assuming an ad hoc finite total time. The algorithm is generic in that any inference technique can be utilized in the E-step. We demonstrate this for exact inference on a discrete maze and Gaussian belief state propagation in continuous stochastic optimal control problems.