StarVLA/Qwen-OFT-Bridge-RT-1

StarVLA QwenOFT for Bridge + RT-1 (10K)

This repository contains the uploaded 10,000-step checkpoint from the 1003_qwenoft run. The public YAML selects the legacy QwenOFT framework and the bridge_rt_1 mixture. Its VLM is stored as an absolute local path ending in Qwen2.5-VL-3B-Instruct; an upstream revision is not recorded.

Model details

Item	Published configuration
Framework	Legacy StarVLA QwenOFT
VLM	Local Qwen2.5-VL-3B-Instruct snapshot; revision missing
Action features	Hidden states at repeated action-placeholder tokens
Action head	Two-block residual MLP, 2,048 input / 4,096 hidden / 7 output
Training objective	Direct L1 action regression
Action dimension / representation	7 / delta_ee
Numeric policy-state input	None in the checkpoint-era QwenOFT action-head path
Action chunk	16 steps
Camera input	image_0, 224 x 224
Dataset mixture	bridge_rt_1
Normalization keys	bridge_dataset, fractal20220817_data
Uploaded checkpoint	checkpoints/steps_10000_pytorch_model.pt

The YAML retains legacy dino, layer_qformer, action_model_type: DiT-B, and diffusion-related fields. The public QwenOFT implementation, including the implementation introduced with this framework, ignores that model-type label and constructs the residual MLP/L1 head described above. No action-head state_dim is declared, while both packaged statistics entries contain 8D raw state. The numeric state is not an MLP input in the matched checkpoint-era implementation. Later opt-in code can discretize state into the instruction, but that is a different prompt contract and should not be enabled for this checkpoint without validation.

Training details

Setting	Value in config.yaml
Intended maximum steps	100,000
Released step	10,000
Per-device VLA batch size	16
Gradient accumulation	1
Warm-up steps	10,000
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 5e-7
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
Seed	42

Evaluation artifacts and incomplete run set

The Hub contains 16 expected log files (four tasks x four runs), but two files terminate before evaluation with vk::PhysicalDevice::createDeviceUnique: ErrorInitializationFailed: Carrot run 3 and Spoon run 3. The remaining log means are:

Task	Completed logs	Mean success over completed logs
PutCarrotOnPlateInScene-v0	3 / 4	0.3333
PutEggplantInBasketScene-v0	4 / 4	0.8750
PutSpoonOnTableClothInScene-v0	3 / 4	0.3750
StackGreenCubeOnYellowCubeBakedTexInScene-v0	4 / 4	0.0833
Completed-log mean	14 / 16	0.4256

The 42.56% completed-log mean is not a complete four-run aggregate and should not be compared directly with complete entries. The StarVLA Model Zoo reports 41.8% for the named model, but does not publish the missing runs needed to reproduce that number from the current Hub artifact.

Files

config.yaml
dataset_statistics.json
checkpoints/
├── steps_10000_pytorch_model.pt
└── steps_10000_pytorch_model_infer_*.log.run{1,2,3,4}

Loading and evaluation

huggingface-cli download StarVLA/Qwen-OFT-Bridge-RT-1 \
  --local-dir Qwen-OFT-Bridge-RT-1

export CKPT=$PWD/Qwen-OFT-Bridge-RT-1/checkpoints/steps_10000_pytorch_model.pt
python deployment/model_server/server_policy.py \
  --ckpt_path "$CKPT" \
  --port 10093 \
  --use_bf16

Use the StarVLA SimplerEnv evaluator with bridge_dataset for WidowX. Re-run the two failed evaluations before publishing a complete aggregate.

Intended use and limitations

This checkpoint is intended for SimplerEnv research with the legacy QwenOFT configuration. The base-model revision is not public and the bundled evaluation set is incomplete. No real-robot or safety validation is claimed. Reconstruct it as the OFT action-token MLP/L1 policy, not from the stale DiT-B configuration label.