StarVLA QwenGR00T Qwen2.5-VL-3B for LIBERO 4-in-1 (30K)
This repository contains the uploaded 30,000-step checkpoint from the
1025_libero_all_qwengroot run. It combines Qwen2.5-VL-3B with the StarVLA
QwenGR00T action framework and was configured on the four-suite
libero_all mixture.
Model details
| Item | Published configuration |
|---|---|
| Framework | StarVLA QwenGR00T |
| VLM | Local snapshot named Qwen2.5-VL-3B-Instruct; revision not recorded |
| YAML-only visual field | dinov2_vits14; the matched QwenGR00T implementation does not construct DINO or Q-Former |
| Action model | 16-layer DiT-B flow head: 768 latent width, 12 heads (64 dimensions/head); state/action decoder MLP width 1,024 |
| Action / model-state dimension | 7 / 7 |
| Raw state statistics | 8D under franka |
| Action representation / chunk | delta_qpos / 8 steps |
| Inference flow steps | 4 |
| Camera input | image_0, 224 x 224 |
| Uploaded checkpoint | checkpoints/steps_30000_pytorch_model.pt |
Use the packaged transform to reconcile the raw 8D state statistics with the 7D action-head state. Camera and action ordering must match the LIBERO adapter. The YAML's DINO field is inherited configuration, not a runtime component of this QwenGR00T path.
Training data
The equal-weight libero_all registry contains the no-noop LeRobot versions
of LIBERO Object, Goal, Spatial, and LIBERO-10. The franka entry in
dataset_statistics.json supplies state/action normalization.
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 |
| Training flow repeats / inference steps | 4 / 4 in the public run-era framework; YAML action-model value 8 is not read there |
| VLA / VLM loss scale | 1.0 / 0.1 |
freeze_modules | Packaged boolean true; the public trainer expects module paths as a string, so this value names/selects no modules |
| Gradient checkpointing / mixed precision | Enabled / enabled |
| Seed | 42 |
summary.jsonl only lists the 10K, 20K, and 30K checkpoint steps.
Evaluation status
Missing for this artifact. No LIBERO evaluation logs or success rates are
present in the repository. The StarVLA paper's 30K StarVLA-GR00T row uses a
Qwen3-VL-4B backbone, whereas this checkpoint is Qwen2.5-VL-3B; that paper row
cannot establish the performance of this artifact.
A future report should use the standard four-suite protocol (10 tasks x 50 episodes per suite), identify the exact checkpoint, and publish per-suite and average success rates.
Files
config.yaml
dataset_statistics.json
summary.jsonl
checkpoints/
└── steps_30000_pytorch_model.pt
Loading and evaluation
huggingface-cli download StarVLA/Qwen2.5-VL-GR00T-LIBERO-4in1 \
--local-dir Qwen2.5-VL-GR00T-LIBERO-4in1
export CKPT=$PWD/Qwen2.5-VL-GR00T-LIBERO-4in1/checkpoints/steps_30000_pytorch_model.pt
python deployment/model_server/server_policy.py \
--ckpt_path "$CKPT" \
--port 10093 \
--use_bf16
Use the maintained
LIBERO evaluation workflow
with normalization key franka.
Intended use and limitations
The checkpoint is intended for LIBERO simulation research. It has one 224 x 224 RGB view, an 8-step 7D delta-joint action contract, no published benchmark result, and no demonstrated real-robot reliability. The exact VLM revision is not pinned in the Hub configuration, and the policy is not safety-tuned.