Reproduce “PISTO: Proximal Inference for Stochastic Trajectory Optimization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hongzhe Yu, Zinuo Chang, Yongxin Chen
The Georgia Institute of Technology researchers developed PISTO (Proximal Inference for Stochastic Trajectory Optimization), an algorithm that unifies STOMP with variational inference and proximal methods to enhance stability and efficiency in robotic motion planning and control. PISTO achieved an 88.57% success rate in motion planning benchmarks, outperforming STOMP (67.59%), and delivered 1.4x to 3.2x faster runtimes with up to 2.8x higher rewards in contact-rich MuJoCo tasks compared to baselines.
Key insights
- STOMP implicitly minimizes the forward KL divergence between a Gaussian proposal and a Boltzmann distribution representing the optimal path posterior.
- Augmenting the forward KL objective with a proximal penalty is equivalent to minimizing the forward KL divergence to a surrogate distribution, which acts as a trust-region mechanism for stability.
- Solving the reverse KL minimization problem with respect to this surrogate distribution allows for derivative-free mean updates using importance sampling, maintaining flexibility for non-differentiable costs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.