junhyeong-vwam-oa-gadapt-cur-humact-agv-50k-g4-bg-1103-v4
Completed 50,000-step EMA inference-only checkpoint. Use step-050000/ only after HF_INFERENCE_COMPLETE.json is present.
EMA tensors are mapped from net_ema.* to the serving net.* names and cast to their matching inference dtype. The original training weights, optimizer, scheduler, trainer and RNG state are excluded. Config and action normalization are included. This is native PyTorch DCP, not a Transformers export.
Use the matching V-WAM runtime and config.yaml, overriding recorded training paths as necessary. Frozen upstream VLM/tokenizer/VAE assets remain runtime dependencies. All tensor names/shapes/dtypes and sampled EMA values are checked on export; remote file sizes and hashes are verified. Robot rollout quality is not evaluated by this backup.