Model Zoo

Available Modified Models

Model	Description	Link
Qwen2.5-VL-3B-Action	Extend Qwen2.5-VL vocabulary with fast tokens (special vocabulary extension for discretizing continuous actions into tokens)	Hugging Face
Qwen3-VL-4B-Action	Extend Qwen3-VL vocabulary with fast tokens (same as above)	Hugging Face
pi-fast	pi-fast action tokenizer weights	Hugging Face

Finetuning Checkpoints

SimplerEnv / Bridge

Bridge is a WidowX tabletop manipulation dataset; Fractal is Google’s RT-1 robot manipulation dataset.

Model	Framework	Base VLM	Description	WidowX	Link
Qwen2.5-FAST-Bridge-RT-1	QwenFast	Qwen2.5-VL-3B	Bridge + Fractal	58.6	HF
Qwen2.5-OFT-Bridge-RT-1	QwenOFT	Qwen2.5-VL-3B	Bridge + Fractal	41.8	HF
Qwen2.5-PI-Bridge-RT-1	QwenPI	Qwen2.5-VL-3B	Bridge + Fractal	62.5	HF
Qwen2.5-GR00T-Bridge-RT-1	QwenGR00T	Qwen2.5-VL-3B	Bridge + Fractal	63.6	HF
Qwen-GR00T-Bridge	QwenGR00T	Qwen2.5-VL-3B	Bridge only	71.4	HF
Qwen3VL-OFT-Bridge-RT-1	QwenOFT	Qwen3-VL-4B	Bridge + Fractal	42.7	HF
Qwen3VL-GR00T-Bridge-RT-1	QwenGR00T	Qwen3-VL-4B	Bridge + Fractal	65.3	HF
Florence-GR00T-Bridge-RT-1	QwenGR00T	Florence-2	Bridge + Fractal (small model)	-	HF

WidowX column: Success rate (%) on WidowX robot tasks in SimplerEnv. Higher is better.

LIBERO

LIBERO has 4 task suites (Spatial, Object, Goal, Long Horizon) with 40 tasks total. All checkpoints are trained jointly on all 4 suites. See LIBERO evaluation docs.

Model	Framework	Base VLM	Link
Qwen2.5-VL-FAST-LIBERO-4in1	QwenFast	Qwen2.5-VL-3B	HF
Qwen2.5-VL-OFT-LIBERO-4in1	QwenOFT	Qwen2.5-VL-3B	HF
Qwen2.5-VL-GR00T-LIBERO-4in1	QwenGR00T	Qwen2.5-VL-3B	HF
Qwen3-VL-OFT-LIBERO-4in1	QwenOFT	Qwen3-VL-4B	HF
Qwen3-VL-PI-LIBERO-4in1	QwenPI	Qwen3-VL-4B	HF

RoboCasa

RoboCasa GR1 Tabletop Tasks with 24 Pick-and-Place tasks. See RoboCasa evaluation docs.

Model	Framework	Base VLM	Link
Qwen3-VL-GR00T-Robocasa-gr1	QwenGR00T	Qwen3-VL-4B	HF
Qwen3-VL-OFT-Robocasa	QwenOFT	Qwen3-VL-4B	HF

RoboTwin

RoboTwin 2.0 dual-arm manipulation benchmark with 50 tasks. See RoboTwin evaluation docs.

Model	Framework	Base VLM	Link
Qwen3-VL-OFT-Robotwin2-All	QwenOFT	Qwen3-VL-4B	HF
Qwen3-VL-OFT-Robotwin2	QwenOFT	Qwen3-VL-4B	HF

BEHAVIOR-1K

BEHAVIOR-1K household task benchmark using R1Pro humanoid robot. See BEHAVIOR evaluation docs.

Model	Description	Link
BEHAVIOR-QwenDual-taskall	Jointly trained on all 50 tasks	HF
BEHAVIOR-QwenDual-task1	Single-task training	HF
BEHAVIOR-QwenDual-task6-40k	6-task joint training	HF
BEHAVIOR-rgp-seg	Segmentation observation experiment	HF

Datasets

Training Datasets

Dataset	Description	Link
LLaVA-OneVision-COCO	Image-text dataset for VLM co-training (ShareGPT4V-COCO subset)	HF
RoboTwin-Clean	RoboTwin 2.0 clean demonstrations (50 per task)	HF
RoboTwin-Randomized	RoboTwin 2.0 randomized demonstrations (500 per task)	HF
RoboTwin-Randomized-targz	Same as above, tar.gz packed format (for bulk download)	HF

BEHAVIOR Data

Dataset	Description	Link
BEHAVIOR-1K	BEHAVIOR-1K benchmark simulation configs	HF
BEHAVIOR-1K-datasets	BEHAVIOR-1K training datasets	HF
BEHAVIOR-1K-datasets-assets	BEHAVIOR-1K scene and object assets	HF
BEHAVIOR-1K-VISUALIZATION-DEMO	BEHAVIOR-1K visualization demos	HF
behavior-1k-task0	Single-task training data sample	HF

How to Use a Checkpoint

Download a checkpoint and run the policy server:

# Download (requires huggingface_hub)
huggingface-cli download StarVLA/Qwen3VL-GR00T-Bridge-RT-1 --local-dir ./results/Checkpoints/Qwen3VL-GR00T-Bridge-RT-1

# Start the policy server
python deployment/model_server/server_policy.py \
    # steps_XXXXX is the training step count — replace with the actual filename from your download
    # e.g. steps_50000_pytorch_model.pt; run `ls` to see the exact filename
    --ckpt_path ./results/Checkpoints/Qwen3VL-GR00T-Bridge-RT-1/checkpoints/steps_XXXXX_pytorch_model.pt \
    --port 5694 \
    --use_bf16

Then follow the evaluation guide for the benchmark you want to test on (e.g. SimplerEnv, LIBERO, RoboCasa, RoboTwin, BEHAVIOR).