bridgequant-so101-76m-h50

BridgeQuant 76M SO101 policy trained with action_horizon: 50 (the openpi / GR00T horizon) instead of 16. Same recipe as bridgequant-so101-76m otherwise, 60k steps.

Run it with n_action_steps >= 25. The demonstrations hold the cube still above the target for up to 49 frames before the final release (median 19). A 16-step chunk contains that release in 29% of the episodes, a 50-step chunk in all of them. Executing only 8 steps per chunk throws the release away again: the gripper opening sits at chunk steps 14 to 39.

Offline, on the stack task (Open the drawer, pick up the red cube then stack on the blue cube): a gripper opening appears in the predicted chunk before the demo's final release in 100% of training episodes (16-step models: about 90%) and in 2 of 5 held-out episodes (16-step models: 0 to 1 of 5). Held-out chunk MAE is unchanged (2.66 vs 2.55).

This checkpoint comes from the hold-out split: the last 5 episodes of each of the 3 datasets were not trained on (154 of 159 episodes used).

Files: model.safetensors, norm_stats.json, bridgequant_vla.yaml / config.json / metadata.json, train_config.yaml (the data paths inside are only needed for training; inference ignores them).

hf download twanghcmut/bridgequant-so101-76m-h50 --local-dir checkpoints/so101-76m-h50

uv run python scripts/so101_policy.py \
    --config checkpoints/so101-76m-h50/train_config.yaml --checkpoint checkpoints/so101-76m-h50 \
    --n-action-steps 25 --self-test \
    --dataset examples/SO101/open_drawer_pick_red_cube_stack_on_blue_cube --n-frames 40
# expected: MAE chunk ~2.2, MAE 8 executed ~1.3, gripper ~1.0 (dataset units)

Code: https://github.com/VLAOpt/bridgequant-async (commit 50683f1 or later) — scripts/so101_policy.py, guide in docs/so101_deploy.md.