TWIST2

Project notes

Validate the simulation route first

This profile reviews commit b06178f19a22f2138cbd31f60c6d494bc263f67d, with no GitHub release listed. Follow the README's two-environment setup: Python 3.8 for controller training/deployment with Isaac Gym, and Python 3.10 for GMR retargeting. Controller setup installs the local rsl_rl, legged_gym and pose packages, Redis and ONNX Runtime, among other dependencies. [1] [6]

From the repository root, the documented first-use sequence initializes the high-level motion server with bash run_motion_server.sh, then starts the simulated low-level controller with bash sim2sim.sh. Keep motion streaming in a separate terminal. The supplied motion launcher selects an example walking file and enables its viewer; the simulation launcher selects the G1 XML and twist2_1017_20k.onnx controller. [1] [2] [3]

Do not use the README's approximate 50 Hz description as a verified operating rate. The reviewed simulation launcher requests --policy_frequency 100 and CUDA. The controller can fall back to ONNX Runtime's CPU provider if its CUDA provider is unavailable, so inspect the reported provider and measured timing for your machine. Neither the configured rate nor upstream sample timing is a benchmark from this review. [1] [2] [4]

The supplied simulation sender and receiver use Redis on the same machine. The README also suggests broad network binding with protected mode disabled; that is a wider exposure than this local playback route needs. Resolve the communication topology and access controls before extending it across machines. [1] [3] [4]

Live teleoperation adds hardware and version dependencies

PICO operation adds the headset client, XRoboToolkit PC Service and a separately built Python binding. The launch script assumes a particular Conda installation path and a fixed human-height value that must match the setup. Laptop-based physical deployment also points to a modified Unitree SDK2 fork, so the official SDK2 alone is not proof of compatibility. [1] [5]

The repository code has MIT terms. Upstream supplies controller and example-motion paths, but this review does not establish separate model/data rights; the larger training dataset is externally linked and AMASS/OMOMO have their own terms. We preserved the source snapshot without running the controller, collecting data or testing physical hardware. [1] [7]