BiGym 2.0
BiGym 2.0: Benchmarking Learned and Agent-Developed Policies for Humanoid Household Manipulation
A benchmark for humanoid household manipulation on Unitree G1 with 20 tasks, 60 VR demos each, evaluating VLA, IL, RL, and coding agents.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info bigym-2
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
5 sourced entriesEvery number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
BiGym 2.0 adapts 20 household tasks for the Unitree G1 robot, providing 60 native human VR demonstrations per task with synchronized multi-camera views. It uses a frozen GR00T-WBC controller for consistent evaluation of learned policies and agent-developed programs.
The benchmark compares vision-language-action fine-tuning, imitation learning, demo-driven RL, and cold-start coding agents. Key findings show agent-developed programs outperform demo-driven RL on bimanual reaching, while cross-workspace stacking remains an open challenge.