snupilab/humanoidtoolbench-cosmos3-real-g1-91
HumanoidToolBench Cosmos3 Edge Policy: Real G1 additional training, 91 demonstrations
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.
Initialization: snupilab/humanoidtoolbench-cosmos3-sim-3003 after 40,000 simulation updates, followed by 5,000 new updates using the 91 real G1 demonstrations.
The four real conditions are StickMove Standard/Reasoning and HookRetrieve Standard/Reasoning. Recordings are 20 Hz. Hardware executed joint-target actions require the matching real G1 control adapter. A simulation adapter must not be assumed compatible.
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.
Native model-only Cosmos3 DCP; optimizer and trainer state excluded. Requires the pinned Cosmos framework and HumanoidToolBench adapter. Run prepare_inference.py with the downloaded repository root, then use the resulting absolute inference-local.json with --checkpoint <root>/model --dataset <root>/dataset --config <root>/inference-local.json. The trained config, raw action units, VAE, and native processor are retained. Native final checks and a fresh CPU model scan passed. The native validator has no historical payload hashes; export hashes identify the bytes checked at publication. GPU reload/evaluation is a separate operation.
Training uses independent model optimizers and shared GPU execution through MPS. Publication is performed by a CPU uploader after final checkpoint validation.