EVALUATION RECORD
PMT RGMT Transformer 20k
LAFAN1 × G1 29DoF · Ideal proprioception · no global position
Evaluated by MotionChallenge using the shared protocol.
Results
23 / 40motions passedCompletion 76.28%
Success and completion use the first declared failure. MPBPE measures position errors over the benchmark’s declared body origins, with three alignments over two time ranges.
| Entire motion | Until first failure | ||||
|---|---|---|---|---|---|
| XY only | XY + yaw | No alignment | XY only | XY + yaw | No alignment |
| 282.57 | 125.24 | 8038.59 | 195.21 | 41.93 | 6966.39 |
Entire-motion errors include every frame through the endpoint, including post-failure states. The other range ends at the first failing sample, including that sample.
- Benchmark
- LAFAN1 × G1 29DoF · Ideal proprioception · no global position
- Engine
- MuJoCo 3.17.0 · official CPU batch
- Checkpoint SHA256
38e8070882d9e397a4850a7b76337fb9260155f6dcc1887759db070f2c014f12- Evaluation code
- MotionChallenge-pmt-rgmt
- Verification
- Both trajectory ranges recomputed · repeatability checked
- Inference comparison
- 160 inputs · 4 resets · Report ↗ · Comparison code ↗
Reproduce this run
git clone --recurse-submodules https://github.com/MotionChallenge-Project/MotionChallenge-pmt-rgmt.git
cd MotionChallenge-pmt-rgmt
uv sync
uv run evaluate.pyRuns the full benchmark and verifies the saved trajectories.
Complete results, source & verification (JSON) ↗