Johns Hopkins University

“Stochastic Multiple Shooting Trajectory Optimization via Sequential Local Policy Evaluation”

“On the cartpole swingup, only multiple shooting achieved convergence to the terminal set, with a mean terminal cost of 23.8 versus 130.3 for MPPI and 123.4 for single-shooting CEM (Table I).”

“MPPI has become 'ubiquitous in the robotics community' (Section II-C) and is widely deployed in autonomous driving, drone racing, and manipulation. Yet this paper shows it completely fails to satisfy terminal constraints on both the cartpole and VTOL landing tasks”

“The authors 'show that we are able to synthesize approximate system Jacobians purely from rollouts, making the method suitable for model-based reinforcement learning with black-box dynamics' (Abstract).”

“The method is validated on a VTOL quadplane with a 22-dimensional state space and 9-dimensional input space—far beyond toy problems. The aircraft performs a precision post-stall landing maneuver, starting from 15 m/s cruise at 15 m altitude, landing 30 m downrange within a 1.2 m radius sphere and sub-2.5 m/s velocity in all axes”

“Co-authored 'Bimanual regrasping for suture needles using reinforcement learning for rapid motion planning' (Reference [3]), which the paper notes can 'directly follow' this needle pickup framework”