STOMP

A gradient-free trajectory optimizer that samples noisy trajectories around an initial guess and updates it by cost-weighted averaging.

STOMP was proposed by Mrinal Kalakrishnan, Sachin Chitta, Evangelos Theodorou, Peter Pastor, and Stefan Schaal at ICRA 2011 as a trajectory optimization method. It starts from an initial trajectory, which may pass straight through obstacles, and each round adds noise around it to generate several candidate trajectories; it computes per-waypoint costs such as collision, constraint violation, and smoothness, then combines the noise into an update weighted so lower-cost trajectories count for more, iterating toward a smooth, collision-free result. STOMP is derived from PI², a path-integral method from reinforcement learning, and it needs no gradient of the cost function, so it can incorporate costs — like torque, energy, or end-effector orientation — that are hard to differentiate. Compared with the gradient-based CHOMP, STOMP's randomness makes it less prone to getting stuck in local optima and needs less tuning; compared with OMPL's sampling-based planners, it is usually slower but produces smoother trajectories that often need no further smoothing afterward.

ExampleMoveIt implements STOMP as a planner plugin. Its behavior is set by parameters such as num_rollouts (how many noisy trajectories to generate per round), stddev (the noise magnitude per joint), and cost functions such as CollisionCheck and ObstacleDistanceGradient.

Also called

Stochastic Trajectory Optimization for Motion Planning

Related

Covariant Hamiltonian Optimization for Motion Planning · TrajOpt · Trajectory Optimization · Motion Planning · MoveIt Motion Planning Framework · Model Predictive Path Integral Control

Sources

MoveIt 2 文档：STOMP Planner (Chinese)
MoveIt 1 教程：STOMP Planner（引用 Kalakrishnan et al. ICRA 2011） (Chinese)