GPU Based Path Integral Control with Learned Dynamics
This work addresses real-time control for robotics applications, such as quadrotor navigation, by enabling efficient planning with learned models, though it is incremental in adapting existing methods.
The paper tackles the problem of real-time optimal control for systems like quadrotors by combining path integral control with learned dynamics models and GPU acceleration, achieving real-time performance in a receding-horizon implementation on a real system.
We present an algorithm which combines recent advances in model based path integral control with machine learning approaches to learning forward dynamics models. We take advantage of the parallel computing power of a GPU to quickly take a massive number of samples from a learned probabilistic dynamics model, which we use to approximate the path integral form of the optimal control. The resulting algorithm runs in a receding-horizon fashion in realtime, and is subject to no restrictive assumptions about costs, constraints, or dynamics. A simple change to the path integral control formulation allows the algorithm to take model uncertainty into account during planning, and we demonstrate its performance on a quadrotor navigation task. In addition to this novel adaptation of path integral control, this is the first time that a receding-horizon implementation of iterative path integral control has been run on a real system.