CRISP - Compliant ROS2 Controllers for Learning-Based Manipulation Policies and Teleoperation

Daniel San José Pro, Oliver Hausdörfer, Ralf Römer, Maximilian Dösch, Martin Schuck, Angela P. Schoellig

Keywords: Foundation Models, Manipulation, Compliance and Impedance Control, Teleoperation, Software-Hardware Integration for Robot Systems, Reinforcement Learning, Imitation Learning, Deployment

TL;DR: A pipeline for learning-based policies and teleoperation: from low-level controllers to Gymnasium interface

Abstract: Learning-based controllers, such as diffusion policies and foundation models output low-frequency and discontinuous target states. Tracking such references smoothly requires a low-level controller to handle such high-level commands and convert them to joint torques, ideally in a compliant way. To this end, we present \textsc{CRISP}, a lightweight C++ implementation of compliant Cartesian and joint-space controllers for the ROS2 control standard designed for seamless integration with learning-based policies and teleoperation. Through our CRISP\_PY Python and CRISP\_GYM Gymnasium interfaces, we provide a unified pipeline to use the CRISP controllers for data collection through teleoperation, policy training, and subsequent policy deployment. The controllers are compatible with any manipulator exposing a joint effort interface and have been validated on numerous platforms, including Franka Robotics FR3 and Universal Robot UR5. Designed for rapid integration, flexible deployment, and real-time performance, our implementation provides a unified pipeline for data collection and policy execution, lowering the barrier to applying learning-based methods on ROS2-compatible manipulators. Detailed documentation is available at https://utiasDSL.github.io/crisp_controllers.

Submission Number: 31