GraphVLA DA3 60k
Research checkpoint from the verified-visible-mask DA3 Edge-GRU run, step 60,000. DA3METRIC-LARGE BF16 geometry, shared-scale stabilization, static-node holding, prompt clique + prompt-only nearest-two union, 5D edge geometry, edge GRU (BPTT=1).
This is an inference export, NOT a resumable training checkpoint. Optimizer and rank memory/RNG states are omitted. No success-rate claim is made for this export. Model code is private: https://github.com/yyyyu120/GraphVLA (access required). The checkpoint requires custom GraphVLA code, Qwen3-VL assets, DA3, fine-tuned YOLOE detector, SAM3, Cutie, MiniLM and the pinned simulator environment. It is not a standalone Transformers AutoModel.
config.yaml preserves the original training paths and compatibility hashes;
deployment must map paths without bypassing the model-contract check.
Action normalization statistics are from the original normalized LIBERO action
targets retained by the independent rerender pipeline.
See SHA256.json for file integrity. Load weights with weights_only=True.
Upstream model and dependency licenses/access conditions continue to apply.
- Downloads last month
- 8