BridgeQuant-Async — LIBERO-Goal 95.0%

Full-finetuned 480M VLA (LFM2.5-VL-450M backbone + BridgeConnector + BridgeActionHead). Best LIBERO-Goal checkpoint of the campaign.

Protocol: seed 42, 200 episodes/suite (20 per task × 10 tasks), synchronous rollout, n_action_steps=8, one Euler denoising step, raw (non-EMA) bundle.

Recipe

train_mode: finetune, lr 5e-5, 30k steps, batch 16, vision 512px, GR00T-style augmentation, taps [0, 6, 12, 15], 64 role queries, memory rank 256, chunk 16, flow matching with Beta(1.5, 1) time sampling, num_steps 1.

Goal-specific: transition-balanced sampling (transition_oversample: 1.5, transition_window: 12) — frames within 12 steps of a debounced gripper open↔close change are sampled 1.5× more often. No auxiliary objective, no architecture change; it only reshapes the data distribution of the same flow loss. Worth +4.5 points on Goal, concentrated in the multi-phase tasks (t06 15→19, t09 16→20).

Async deployment

The same weights run under the asynchronous runtime (cached bridge latent, ~5 ms replan) at 94.0 with one fixed controller — K=8, --context-refresh-every 2, --precision-refresh --precision-refresh-approach --precision-refresh-uncertainty — no per-suite tuning.

Files

model.safetensors (trainable weights incl. the finetuned backbone), norm_stats.json, config.json, bridgequant_vla.yaml (config saved with the checkpoint), train_config.yaml (the exact training YAML).

Eval

uv run python scripts/eval_libero_object.py \
  --config train_config.yaml --checkpoint <this-dir> \
  --n-action-steps 8 --episodes-per-task 20

The eval suite must match the training suite or success rate is 0%.