PISTO: New robot motion planner achieves 89% success rate, 2x faster

A team of researchers led by Hongzhe Yu has unveiled PISTO (Proximal Inference for Stochastic Trajectory Optimization), a new algorithm that significantly advances robot motion planning. The key insight: existing stochastic trajectory optimization methods like STOMP implicitly minimize KL divergence from a Boltzmann distribution—revealing an elegant variational inference structure. PISTO builds on this by adding a proximal KL regularization term between successive Gaussian proposals, which stabilizes updates and enables a trust-region-like behavior. The result is a closed-form mean update computable via importance-weighted Monte Carlo sampling, making the algorithm fully derivative-free and capable of handling non-differentiable or even discontinuous cost functions.

In rigorous benchmarks on robot arm motion planning, PISTO achieved an 89% success rate, outperforming CHOMP (63%) and STOMP (68%) by wide margins. It also produced shorter, smoother trajectories at roughly twice the computational speed of competing stochastic methods. Additional validation on contact-rich MuJoCo locomotion and manipulation tasks showed consistent superiority over CEM and MPPI baselines in total reward. This work not only provides a practical, high-performance tool for robotics but also deepens theoretical understanding by connecting stochastic trajectory optimization to variational inference.