JAX-Powered MPC Outperforms Baselines in Real-World Robot Contact Tasks

Researchers from TU Darmstadt have demonstrated the real-world deployment of a massively parallel sampling-based Model Predictive Control (MPC) framework for contact-rich robotic manipulation. The system leverages JAX for large-scale parallelization and the high-fidelity MuJoCo MJX simulator, deployed on a Franka Research 3 arm executing the Push-T task. Their approach uses a real-to-sim-to-real pipeline, enabling the robot to learn and adapt in simulation before transferring directly to hardware.

The key innovation is the MTP (Mixture of Trajectory Priors) variant with structured global sampling, which significantly outperformed unimodal baselines like CEM, MPPI, and PS in tasks requiring mode switching—both in simulation and on the real robot. The team also evaluated online domain randomization within the MPC sample budget, finding that contact-initiation parameters yield interpretable adaptation signals, whereas global physics parameters provide feedback too weak for reliable exploitation at typical replanning frequencies.

These results highlight critical challenges for sampling-based MPC in contact-rich manipulation: contact sensitivity, tight compute budgets, and the difficulty of obtaining informative domain-randomization signals in real time. The work represents a step toward robust, transferable control for robots operating in unstructured environments.