HumanoidToolBench
Collection
HumanoidToolBench simulation policies (3,003 segments), selected real G1 adaptations (91 demos), and training data. • 15 items • Updated
HumanoidToolBench pi0.5: Simulation training, 3,003 segments
The validated final policy checkpoint is available in this repository.
This is a HumanoidToolBench training-result repository. It does not substitute an upstream pretrained policy for a HumanoidToolBench-trained checkpoint.
| Setting | Value |
|---|---|
| Training stage | Simulation training, 3,003 segments |
| Target optimizer updates | 40000 |
| Per-GPU batch / GPUs / global batch | 16 / 8 / 128 |
| Gradient accumulation | 1 |
| Conditions per global batch | 18 |
| Dataset revision | 8b2cd31e107b64cb13f812ea217a63a20845c78a |
The simulation pool contains 1,200 successful L1/L2 demonstrations and 1,803 extracted L0 prefixes, spanning 18 conditions. The 3,003 segments are not 3,003 independent demonstrations.
Use the model's native HumanoidToolBench adapter and model-specific dependencies. This repository does not claim compatibility with arbitrary Transformers or simulation loaders. No evaluation score is claimed by checkpoint publication.
Use pinned theta_bench.cli._openpi_runner directly from the exported repository root: parse ['serve', *launch_json['runner_args'], '--checkpoint=/run/40000', '--host=127.0.0.1', '--port=8000'] with _parser(), then call serve(args). Set DATA_HOME to the exported data directory and PSI_HOME to the pinned native Psi0 source. The outer pi05-humanoid CLI requires the original bootstrap checkpoint and changes cwd, so its unmodified relative-path wrapper is not the portable export entry point. Set OPENPI_DATA_HOME to the absolute exported openpi_cache path so the native Paligemma tokenizer uses the bundled exact bytes. The checkpoint-local norm_stats and asset ID are preserved. Retain pi05=True, 36 action slots, H16, 250 language/state tokens and masked-image-mode=encoder; the native HFM image transform defines missing wrist slots and masks. Base policy weights, optimizer, RNG and metadata.pt are unnecessary for inference.
Training uses independent model optimizers and shared GPU execution through MPS. Publication is performed by a CPU uploader after final checkpoint validation.