Robustness without Wrinkles: Parallel Simulation and Robust MPC for Certified Deformable Manipulation
Researchers have developed CORD-SLS, a novel real-time control method for the safe manipulation of deformable objects like ropes and cloth. This method utilizes a GPU-parallel differentiable simulator for efficient gradient-based planning and a robust model predictive control (MPC) algorithm to handle uncertainty. CORD-SLS also incorporates conformal prediction to calibrate visual feedback and perception error bounds, enabling high-probability safe control and accelerating model-based reinforcement learning for training manipulation policies. Evaluations in simulation and hardware demonstrate CORD-SLS's ability to achieve millisecond-speed planning and outperform baseline methods in safety, speed, and task success. AI
IMPACT This research could advance robotics by enabling more precise and safe manipulation of flexible materials, potentially impacting fields like manufacturing and logistics.