bridgequant-so101-76m

BridgeQuant 76M (DINOv3 ViT-S + BERT-mini, 4-layer fusion trunk, role-query bridge, 6-layer flow expert) trained on three SO101 follower-arm tasks (159 episodes, 30 fps, cameras top+wrist, 6-D absolute joint targets). In-sample MAE 1.31 (16-step chunk) / 1.11 (8 executed steps).

Files: model.safetensors (trainable weights), norm_stats.json (state/action normalisation), bridgequant_vla.yaml / config.json / metadata.json (the bundle the loader expects), train_config.yaml (the experiment YAML used for training; evaluation and deployment take the same YAML).

Code: https://github.com/VLAOpt/bridgequant-async — load with load_trainable_weights(policy, "<this dir>") after BridgeQuantAsyncVLA.from_pretrained(config=cfg.policy, init_empty_trainable=True); for the SO101 arm use scripts/so101_policy.py (see docs/so101_deploy.md).