HumanoidToolBench
Collection
HumanoidToolBench simulation policies (3,003 segments), selected real G1 adaptations (91 demos), and training data. • 3 items • Updated
HumanoidToolBench ACT100, SIM 3003
Final language-conditioned native ACT100 checkpoint after 40,000 optimizer updates on 3,003 simulation segments. Full model, optimizer, scheduler, eight-rank RNG and sampler validation passed before inference export. The checkpoint guardian completed validation and retention.
The model predicts 100 actions at 50 Hz (2 seconds). Saved learned query shape is 100 x 512 and VAE position shape is 1 x 102 x 512. ACT16 weights are incompatible.
Training uses batch 16 per GPU on eight NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs (96 GB each), accumulation 1 and seed 1000. Common H40 eligibility yields 2,205,328 anchors across 18 conditions. Padded temporal targets remain masked.
Evaluate it with the HumanoidToolBench toolkit: uv run htb-eval --model snupilab/humanoidtoolbench-act-sim-3003. Exact preprocessing, bounds statistics and instructions accompany the weights. This is a native HumanoidToolBench/Psi0 release, not a Transformers AutoModel package.
Simulation uses the decoupled_v1 action contract.
This release reports training and checkpoint validation only. No simulator evaluation score, validated rollout video or hardware deployment result is claimed. Historical ACT16 Git revisions remain historical; the current release replaces their withdrawn weights.
Data revision: 8b2cd31e107b64cb13f812ea217a63a20845c78a. Source revision: e2352386296d4678d852f79c1d174400f87e1242. See training_provenance.json and validation_summary.json for provenance.