bridgequant-async-libero-long-94

BridgeQuant-Async — LIBERO-Long (LIBERO-10) 94.0%

Full-finetuned 480M VLA (LFM2.5-VL-450M backbone + BridgeConnector + BridgeActionHead). Best LIBERO-Long checkpoint of the campaign — within 0.35 points of GR00T N1.7 at 1/4 the parameters and 1/26 the sample passes.

Protocol: seed 42, 200 episodes, synchronous rollout, n_action_steps=16, one Euler denoising step, raw (non-EMA) bundle.

Why K=16 here

Long is monotone in commitment: K=1/4/8/16 gives 54.5/80.0/89.0/94.0. Each replan re-draws from a multimodal flow sampler, so replanning mid-sequence lets the policy switch targets — on "put both moka pots on the stove" (two identical objects) that costs 7 of 20 episodes. Committing a full 16-step chunk fixes it. n_action_steps is a runtime knob: the loss is uniform over the trained chunk.

Recipe

Plain full finetune: train_mode: finetune, lr 5e-5, 30k steps, batch 16, vision 512px, GR00T-style augmentation, taps [0, 6, 12, 15], chunk 16, flow matching with Beta(1.5, 1) time sampling, num_steps 1. No auxiliary language objective, no transition oversampling — on this suite every extra shaping signal measured so far destroys one rare skill ("turn on the stove, then place the pot": 19/20 at this recipe, ~0 under all ten perturbations tried).

Async deployment

Same weights, asynchronous runtime (~5 ms replan): 93.0 at K=16 with RTC-freeze 4 + uncertainty refresh, or 90.0 under the unified K=8 controller — which is better than synchronous K=8 (89.0), because the controller refreshes vision at grasp and on sampler disagreement rather than on a fixed 8-step clock.

Eval

uv run python scripts/eval_libero_object.py \
  --config train_config.yaml --checkpoint <this-dir> \
  --n-action-steps 16 --episodes-per-task 20