Every night, we train a policy to imitate a reference motion on the Unitree G1 using the latest MJLab commit. The trained policy is then evaluated across 1024 trials, with the results logged here to catch regressions over time.
| Date | Commit | Run | Success Rate | MPKPE (m) | EE Pos Error (m) |
|---|
Measures simulation throughput in env steps per second for different tasks. Physics SPS is the theoretical max if only physics ran (no managers). Env SPS is the actual throughput including observations, rewards, terminations, and resets. Overhead is the percentage of time spent on non-physics work. Benchmarks run on NVIDIA RTX 5090 with 4096 parallel environments.
Note: These benchmarks use zero actions, which causes frequent terminations and resets. This represents a worst-case scenario for overhead since resets are expensive.
| Date | Commit | Task | Physics SPS | Env SPS | Overhead |
|---|