17/09/2025
Looking for an easy-to-use way to deploy your VLAs and diffusion policies on real robots (maybe even before the ICRA video deadline next week 👀)?
🚀 Excited to share our latest work: 𝗖𝗥𝗜𝗦𝗣 - 𝗖𝗼𝗺𝗽𝗹𝗶𝗮𝗻𝘁 𝗥𝗢𝗦2 𝗖𝗼𝗻𝘁𝗿𝗼𝗹𝗹𝗲𝗿𝘀 𝗳𝗼𝗿 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴-𝗕𝗮𝘀𝗲𝗱 𝗠𝗮𝗻𝗶𝗽𝘂𝗹𝗮𝘁𝗶𝗼𝗻 𝗣𝗼𝗹𝗶𝗰𝗶𝗲𝘀 𝗮𝗻𝗱 𝗧𝗲𝗹𝗲𝗼𝗽𝗲𝗿𝗮𝘁𝗶𝗼𝗻!
Why CRISP?
Most learning-based policies (diffusion, VLAs, etc.) output low-frequency or discontinuous commands, which don’t play nicely with real hardware. CRISP bridges that gap with lightweight, compliant, torque-based controllers built on hashtag control.
Features:
🐍 𝗣𝘆𝘁𝗵𝗼𝗻 𝗶𝗻𝘁𝗲𝗿𝗳𝗮𝗰𝗲 — control your robot around without worrying about topics, spinning, etc., but still keep full ROS2 flexibility.
🔁 𝗚𝘆𝗺𝗻𝗮𝘀𝗶𝘂𝗺 𝗲𝗻𝘃𝗶𝗿𝗼𝗻𝗺𝗲𝗻𝘁 — use teleoperation to record robot and visual data in hashtag format & seamlessly deploy learning-based policies.
🎥 𝗗𝗲𝗺𝗼-𝗿𝗲𝗮𝗱𝘆 — examples for single-arm & bimanual manipulation with the Franka FR3.
🤖 𝗖𝗮𝗿𝘁𝗲𝘀𝗶𝗮𝗻 & 𝗷𝗼𝗶𝗻𝘁-𝘀𝗽𝗮𝗰𝗲 𝗰𝗼𝗻𝘁𝗿𝗼𝗹𝗹𝗲𝗿𝘀 – built for torque control and smooth, compliant interaction.
🚫 𝗡𝗼 𝗠𝗼𝘃𝗲𝗜𝘁 𝗼𝗿 𝗵𝗲𝗮𝘃𝘆 𝗽𝗹𝗮𝗻𝗻𝗶𝗻𝗴 𝘀𝘁𝗮𝗰𝗸𝘀 required, ready to use.
Our goal: Make deploying learning-based methods on real robots as frictionless as possible, reducing the gap between data collection, simulation, and deployment.
📄 Paper: https://arxiv.org/abs/2509.06819
💻 Code: https://github.com/utiasDSL/crisp_controllers
🌐 Website: https://utiasdsl.github.io/crisp_controllers/
This work is the result of a great team effort by Daniel San José Pro, Oliver Hausdörfer, Ralf Römer, Maximilian Dösch, and Martin Schuck!
