You can deploy this model and start earning money today!

twanghcmut/
bridgequant-async-libero-unified-95

Name: bridgequant-async-libero-unified-95
Availability: InStock

safetensors

robotics

vision-language-action

Views

No views yet

BridgeQuant-Async — Unified Recipe (LIBERO, mean 95.0 sync / 93.5 async)

One architecture, one training recipe, one runtime controller — no per-suite tuning anywhere. Four checkpoints (one per LIBERO suite), all trained with the identical recipe; the only difference between folders is the training data (the suite's own demos, original demos only).

Backbone: LFM2.5-VL-450M (full finetune) → BridgeConnector → BridgeActionHead (flow matching, chunk 16). ~480M params total.
Recipe (identical for all four): lr 5e-5, 30k steps, batch 16, seed 42, smoothness 0.05, transition oversample 1.25/12, language-NCE 0.01, adaptive task balancing α=0.25 (GroupDRO-lite: per-task weight ∝ (EMA task loss / mean)^0.25, clipped 4×, mass-conserving, grouped by instruction string — no task identity is hard-coded).
Runtime protocol: n_action_steps K=8, seed 42, 200 episodes/suite. Async controller: refresh 2 + precision-refresh(+approach) + uncertainty refresh, no RTC.

Results (seed 42, 200 ep/suite)

mode	Object	Spatial	Goal	Long	mean
sync K=8	98.5	96.0	96.5	89.0	95.0
async K=8	98.0	95.0	95.5	85.5	93.5

Goal 96.5 is the highest goal number we have measured under any recipe, including goal-specific ones. For per-suite-tuned checkpoints (mean 96.75) see the twanghcmut/bridgequant-async-libero-{object-100,spatial-98,goal-95,long-94} repos.

Files

Each suite folder (object/, spatial/, goal/, long/) is a self-contained eval-ready bundle:

model.safetensors — trainable weights (connector + action head + finetuned backbone deltas, exported by requires_grad)
norm_stats.json — action/state normalization (mean/std)
config.json, metadata.json — bundle metadata
bridgequant_vla.yaml / train_config.yaml — the exact training config (configs are self-contained: policy + eval + train)

Usage

bash

git clone https://github.com/<repo>/bridgequant-async  # project code
uv run python scripts/eval_libero_object.py \
    --config <suite>/train_config.yaml \
    --checkpoint <suite>/ \
    --n-action-steps 8 \
    --output-dir eval_outputs/unified_<suite>

The eval suite must match the folder's training suite. chunk_size=16 is baked into the weights; K (--n-action-steps) is freely tunable ≤ 16.

twanghcmut/bridgequant-async-libero-unified-95

BridgeQuant-Async — Unified Recipe (LIBERO, mean 95.0 sync / 93.5 async)

Results (seed 42, 200 ep/suite)

Files

Usage

twanghcmut/
bridgequant-async-libero-unified-95