StarVLA QwenOFT Qwen2.5-VL-3B for LIBERO 4-in-1 (30K)
This repository publishes one 30,000-step checkpoint from the
1025_libero_all_qwengoft run. The trainer target in config.yaml is 80K and
summary.jsonl lists checkpoints through 50K, but no 40K or 50K weight file is
present; the model available to download is the 30K artifact.
Model details
| Item | Published configuration |
|---|---|
| Framework | StarVLA QwenOFT |
| VLM | Local snapshot named Qwen2.5-VL-3B-Instruct; revision not recorded |
| Action features | Hidden states at repeated action-placeholder tokens |
| Action head | Two-block residual MLP with direct L1 regression |
| Action dimension | 7 |
| Numeric policy-state input | None in the checkpoint-era QwenOFT action-head path; YAML state_dim: 7 is not an MLP input |
| Packaged raw state statistics | 8D under franka; not evidence of a direct policy-state input |
| 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 QwenOFT implementation does not build
those components: its action-model factory constructs a residual MLP, consumes
VLM action-placeholder hidden states, and optimizes L1 action loss. Later
opt-in code can serialize state into the instruction, but no numeric state is
fed to this checkpoint's MLP head. The stale fields must therefore not be
presented as runtime architecture.
Training data
libero_all combines equal-weight, no-noop LeRobot versions of LIBERO Object,
Goal, Spatial, and LIBERO-10. The packaged franka statistics provide the
action normalization contract; their state entry comes from the dataset and
does not imply direct state conditioning for this policy.
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 |
Evaluation status
Missing for this artifact. There are no LIBERO rollout logs or quantitative results in the Hub repository. The Qwen3-VL-4B OFT numbers in the StarVLA paper refer to a different backbone/release and are not assigned to this Qwen2.5-VL checkpoint.
The expected reproducible report is 50 episodes for each of 10 tasks in every suite, with Object, Goal, Spatial, LIBERO-10, and their unweighted average reported separately.
Files
config.yaml
dataset_statistics.json
summary.jsonl
checkpoints/
└── steps_30000_pytorch_model.pt
Loading and evaluation
huggingface-cli download StarVLA/Qwen2.5-VL-OFT-LIBERO-4in1 \
--local-dir Qwen2.5-VL-OFT-LIBERO-4in1
export CKPT=$PWD/Qwen2.5-VL-OFT-LIBERO-4in1/checkpoints/steps_30000_pytorch_model.pt
python deployment/model_server/server_policy.py \
--ckpt_path "$CKPT" \
--port 10093 \
--use_bf16
Use the
StarVLA LIBERO guide
and normalization key franka for the simulator client.
Intended use and limitations
This is a LIBERO simulation research checkpoint. Its exact VLM revision, benchmark score, and historical code revision are not fully recorded. The public OFT implementation establishes the MLP/L1 action-head structure, while the stale YAML fields should not be used to instantiate a DiT. It has not been validated for other embodiments or physical robots and is not safety-tuned.