StarVLA QwenFAST Qwen2.5-VL-3B for LIBERO 4-in-1 (30K)
This repository contains the uploaded 30,000-step checkpoint from the
1025_libero4in1_qwenfast run. It uses the StarVLA QwenFast framework and
the libero_all mixture, which combines LIBERO Object, Goal, Spatial, and
LIBERO-10 with equal registry weights. The policy tokenizes continuous action
chunks with FAST and predicts the resulting action tokens autoregressively with
the VLM; it does not run a diffusion sampler.
Model details
| Item | Published configuration |
|---|---|
| Framework | StarVLA QwenFast |
| VLM | Local snapshot named Qwen2.5-VL-3B-Instruct-Action; revision not recorded |
| Action encoding | FAST tokenizer over continuous action chunks |
| Training objective | Autoregressive cross-entropy over mapped action tokens |
| Inference | VLM token generation followed by FAST-token decoding |
| Action dimension | 7 |
| Numeric policy-state input | None in the checked QwenFast forward and inference paths; YAML state_dim: 7 is unused there |
| Packaged raw state statistics | 8D under franka; dataset artifact only for this policy path |
| Action representation / chunk | delta_qpos / 8 steps |
| Camera input | image_0, 224 x 224 |
| Uploaded checkpoint | checkpoints/steps_30000_pytorch_model.pt |
The legacy YAML retains dino, action_model_type: DiT-B, diffusion, and
state_dim: 7 fields. The selected QwenFast implementation does not
construct a DINO/DiT head or read numeric state: it trains the VLM on FAST
action tokens. Do not turn the packaged 8D dataset state statistics into a
model input that the checkpoint was not trained to consume.
Training data
libero_all resolves to four LeRobot datasets:
libero_object_no_noops_1.0.0_lerobotlibero_goal_no_noops_1.0.0_lerobotlibero_spatial_no_noops_1.0.0_lerobotlibero_10_no_noops_1.0.0_lerobot
The packaged dataset_statistics.json contains the franka normalization
entry used at inference.
Training details
| Setting | Value in config.yaml |
|---|---|
| Intended maximum steps | 80,000 |
| Released step | 30,000 |
| Per-device VLA batch size | 16 |
| Gradient accumulation | 1 |
| Warm-up steps | 5,000 |
| Base / interface / action LR | 3e-5 / 1e-5 / 1e-4 |
| Optimizer | AdamW, betas (0.9, 0.95), epsilon 1e-8 |
| Scheduler | Cosine with minimum LR 1e-6 |
| VLA / VLM loss scale | 1.0 / 0.1 |
Packaged freeze_modules | Boolean true; the public trainer only accepts a comma-separated string, so this does not select modules to freeze |
| Gradient checkpointing / mixed precision | Enabled / enabled |
| Seed | 42 |
summary.jsonl lists 10K, 20K, and 30K. It does not contain success rates.
Evaluation status
Missing for this artifact. The Hub repository contains no LIBERO rollout logs, per-suite success table, seed list, or aggregate result. The StarVLA paper reports 30K LIBERO results for Qwen3-VL-4B variants, but this checkpoint uses a Qwen2.5-VL-3B backbone, so those values are not transferred to this card.
For a future comparable report, the project protocol evaluates all 10 tasks in each suite with 50 episodes per task (500 trials per suite) and should record the evaluator revision and seed.
Files
config.yaml
dataset_statistics.json
summary.jsonl
checkpoints/
└── steps_30000_pytorch_model.pt
Loading and evaluation
huggingface-cli download StarVLA/Qwen2.5-VL-FAST-LIBERO-4in1 \
--local-dir Qwen2.5-VL-FAST-LIBERO-4in1
export CKPT=$PWD/Qwen2.5-VL-FAST-LIBERO-4in1/checkpoints/steps_30000_pytorch_model.pt
python deployment/model_server/server_policy.py \
--ckpt_path "$CKPT" \
--port 10093 \
--use_bf16
Run the simulator in a separate environment using the
StarVLA LIBERO guide
and select normalization key franka.
Intended use and limitations
This checkpoint is intended for LIBERO simulation research with a single RGB
view and 7D delta-joint-position actions. Its runtime action representation is
autoregressive FAST token generation, even though the legacy YAML labels an
unused action-model block DiT-B. No public score, physical-robot validation,
or safety guarantee is provided.