// the find
Improbable-AI/walk-these-ways
Sim-to-real RL training and deployment tools for the Unitree Go1 robot.
Research code from MIT's Improbable AI lab implementing the 'Walk These Ways' CoRL 2022 paper — sim-to-real RL locomotion for the Unitree Go1 quadruped. Trains a gait-conditioned policy in Isaac Gym with Multiplicity of Behavior (MoB), then deploys it directly on the physical robot via LCM. Target audience is robotics researchers with a Go1 Edu and a beefy NVIDIA GPU.
Ships a pretrained checkpoint you can deploy on day one without training anything — rare for academic robotics code. The sim-to-real pipeline is end-to-end: Isaac Gym training, domain randomization, actuator net to model motor dynamics, Docker deployment image on the robot's Jetson. The MoB approach of conditioning on gait parameters (frequency, stance duration, foot height) is a real idea that generalizes better than single-policy controllers. LCM for real-time comms between the onboard computer and the policy inference loop is the right tool here — low latency, typed messages.
Pinned to PyTorch 1.10 and CUDA 11.3, which are four years old. Isaac Gym Preview 4 is itself deprecated by NVIDIA in favor of Isaac Lab, so the simulation dependency is a dead end — anyone starting a new project today would be building on a sunset platform. The deployment docs have two 'Coming soon' sections (logging and real-world analysis), which are arguably the most important parts when something goes wrong on a real robot. Swapping in a custom model requires manually editing hardcoded path strings in two separate scripts rather than taking a CLI argument — minor but annoying.