Qwen3-VL PI for LIBERO 4-in-1
This repository contains a StarVLA QwenPI checkpoint initialized from
Qwen3-VL-4B-Instruct and trained on the combined LIBERO
libero_10 + libero_goal + libero_object + libero_spatial mixture. The model
uses Qwen3-VL features with a layer-wise cross-DiT flow-matching action head.
Model and action contract
| Item | Value |
|---|---|
| Framework | StarVLA QwenPI |
| Base VLM | Qwen3-VL-4B-Instruct |
| Action head | Post-restructure layer-wise cross-DiT: 36 layers, 2,560 latent width, 40 heads; each block receives the corresponding one of the last 36 Qwen3-VL hidden layers |
| Action representation | 7D delta_qpos |
| Configured state dimension | 7 |
| Action horizon | 8 |
| Camera input | One image_0 RGB view, default 224 × 224 |
| Inference flow steps | 4 |
| Dataset mixture | libero_all |
| Released checkpoint | checkpoints/steps_100000_pytorch_model.pt |
Training recipe
| Setting | Value |
|---|---|
| Maximum and released step | 100,000 |
| VLA per-device batch | 8 |
| Auxiliary VLM per-device batch | 4 |
| Gradient accumulation | 1 |
| Warm-up | 5,000 steps |
| Base / VLM-interface / action LR | 3e-5 / 1e-5 / 1e-4 |
| VLA / auxiliary VLM loss scale | 1.0 / 0.1 |
| Training flow repeats | 2 in the matching public QwenPI implementation; YAML action-model value 8 is overridden |
freeze_modules | Packaged boolean true; the public trainer expects module paths as a string, so this value names/selects no modules |
| Seed | 42 |
LIBERO evaluation
The following success rates are retained from the original Hub Card.
| Training Steps | libero_goal | libero_object | libero_spatial | libero_10 | Average Score |
|---|---|---|---|---|---|
| 10,000 | 0.752 | 0.850 | 0.896 | 0.460 | 0.740 |
| 20,000 | 0.834 | 0.886 | 0.952 | 0.636 | 0.827 |
| 30,000 | 0.958 | 0.996 | 0.988 | 0.884 | 0.957 |
| 40,000 | 0.948 | 0.982 | 0.986 | 0.860 | 0.944 |
| 50,000 | 0.974 | 0.998 | 0.984 | 0.928 | 0.971 |
| 60,000 | 0.974 | 0.996 | 0.976 | 0.908 | 0.964 |
| 70,000 | 0.974 | 0.992 | 0.986 | 0.930 | 0.971 |
| 80,000 | 0.972 | 0.976 | 0.990 | 0.958 | 0.974 |
| 90,000 | 0.970 | 0.990 | 0.976 | 0.952 | 0.972 |
| 100,000 | 0.962 | 0.986 | 0.992 | 0.960 | 0.975 |
Only the 100k weight is included. The repository does not include raw per-episode evaluation logs for the intermediate rows.
Download and evaluation
This released checkpoint uses the historical, non-canonical LayerwiseFM forward behavior. Use the documented compatibility override:
huggingface-cli download StarVLA/Qwen3-VL-PI-LIBERO-4in1 \
--local-dir Qwen3-VL-PI-LIBERO-4in1
CKPT="$PWD/Qwen3-VL-PI-LIBERO-4in1/checkpoints/steps_100000_pytorch_model.pt"
python deployment/model_server/server_policy.py \
--ckpt_path "$CKPT" --port 6694 --use_bf16 \
--config_override framework.qwenvl.base_vlm=Qwen/Qwen3-VL-4B-Instruct \
--config_override framework.action_model.diffusion_model_cfg.use_canonical_forward=false
Run the simulator separately according to StarVLA's LIBERO guide.
Evidence and limitations
- Structure, action contract, and training values were checked against
config.yaml; checkpoint availability was checked against the Hub file tree. - The YAML's
dino,action_model_type: DiT-B, 16-layer, and repeat-8 fields are not the constructed post-restructure QwenPI shape. The matching public implementation derives the 36-layer, 2,560-wide head from Qwen3-VL and does not construct DINO. - Evaluation numbers are Card-reported. The Hub repository does not package the raw rollout logs needed to independently aggregate them.
- The saved base-model and data paths are environment-local and require mapping.
- Results depend on the historical LayerwiseFM compatibility override, normalization statistics, camera/state contract, and evaluator version.
- Results cover LIBERO simulation only; real-robot and cross-embodiment performance are not established.