lixinyao11/contactmimic

ContactMimic is a research project for making a humanoid robot interact realistically with objects, like sitting in a chair, by training it to reproduce not just human motion but also exactly where and when the human's body touches an object. It comes from an academic paper of the same name and is built specifically for the Unitree G1 humanoid robot inside NVIDIA's Isaac Lab physics simulator. The core idea is that most motion imitation systems only teach a robot to copy body poses over time, without paying attention to contact, the points where hands, feet, or other body parts actually press against an object's surface. ContactMimic adds contact information as part of what the robot learns to reproduce, so a simulated humanoid can learn to sit down and match the same seat, backrest, and leg contact patterns a real human used, rather than just approximating the pose. Training data comes from the HUMOTO motion capture dataset, which requires signing a separate license agreement before you can access the raw recordings. Once you have that raw data, you still need to retarget it onto the G1 robot's body and add labels marking, frame by frame, which parts of the robot and object are touching. The project includes a conversion script that turns this retargeted, contact labeled data into a format the training pipeline can use directly, resampling everything to a consistent frame rate. To set the project up, you install a specific version of Isaac Lab, clone this repository, download the G1 robot's body description files, and install the project as a Python package inside your Isaac Lab environment. Training is run through a provided script that lets you choose which motions to train on and offers several contact specific options, such as turning contact information off entirely for comparison, randomly swapping contact labels between paired motions, or randomly masking out certain contact parts to make the resulting behavior more robust. A separate evaluation script lets you load a trained run and watch, or record video of, the resulting robot behavior in simulation. The project builds on an earlier system called BeyondMimic and is released under the MIT license, so it can be used, modified, and shared freely, though the motion capture data itself is governed by HUMOTO's own license terms.