From Simulation to Contact: A Modular Wrench-Space Deployment Stack for Multi-Task Robot Manipulation

Aditya Narendra, Abdalla Swikir

Keywords: multi-task reinforcement learning, sim-to-real transfer, compliant control, contact-rich manipulation, wrench-space control

Abstract: Multi-task reinforcement learning has shown strong performance in simulation, but transferring such policies to real manipulators remains difficult when low-rate policy outputs interact with high-frequency contact dynamics. We propose MoCE, a compliant high-frequency multi-task reinforcement-learning framework for real robot manipulation. MoCE combines a simulation-trained multi-task policy based on M3PO with a high-rate compliant execution layer implemented with CRISP on a Franka arm. The policy operates on state observations only and outputs a 6D end-effector wrench together with a gripper command, allowing a single shared action space across tasks. The low-rate learned action is executed through a high-frequency compliant controller, enabling stable contact-rich behavior during real-world deployment. We evaluate the method on two tasks, insertion and stacking, in both simulation and on hardware. Our study reports simulated multi-task success rates for each task and their average, and compares simulation and real deployment through compute time, inference time, torque, and joint-velocity metrics. The goal of this paper is to show that compliant high-frequency control is a practical bridge from multi-task RL in simulation to contact-rich execution on a real Franka robot.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 22