StarVLA/Qwen-PI-Bridge-RT-1

StarVLA QwenPI for Bridge + RT-1 (30K)

This repository contains the uploaded 30,000-step checkpoint from the 1011_qwenpi run. The configuration uses the legacy StarVLA QwenPI framework (not QwenPI_v3) with the action-extended Qwen2.5-VL backbone and the bridge_rt_1 training mixture.

Model details

Item	Published configuration
Framework	StarVLA QwenPI
Base VLM	StarVLA/Qwen2.5-VL-3B-Instruct-Action
YAML-only visual field	dinov2_vits14; the matched legacy QwenPI implementation does not construct DINO or Q-Former
Action model	Pre-2025-10-20 layer-wise cross-DiT flow head: 16 layers, 2,048 latent width, 32 heads; conditioned by the last 16 VLM hidden layers
Action / configured head-state dimension	7 / 7
Raw state statistics	8D for oxe_bridge and oxe_rt1
Action representation / chunk	delta_ee / 16 steps
Inference flow steps	4
Camera input	image_0, 224 x 224
Uploaded checkpoint	checkpoints/steps_30000_pytorch_model.pt

The YAML's DiT-Qwen selector supplies the 2,048-wide template, while diffusion_model_cfg.num_layers: 16 overrides its template depth. The public framework contemporary with this legacy checkpoint consumes layer-wise VLM features and does not construct the YAML's DINO/Q-Former fields. Use the packaged transform when state is supplied so the raw 8D statistics are not fed directly to the configured 7D state encoder.

Training details

Setting	Value in config.yaml
Intended maximum steps	100,000
Released and evaluated step	30,000
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
Training flow repeats	Not recoverable exactly: YAML stores action-model 8 and trainer 4, while public legacy code hard-coded 1 on October 12 and 2 on October 14; the run has no code revision
Inference flow steps	4
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
Seed	42

SimplerEnv WidowX evaluation

All 16 uploaded task/run logs complete and report an Average success value. The following table averages four runs for each task.

Task	Success rate
PutCarrotOnPlateInScene-v0	0.6042
PutEggplantInBasketScene-v0	0.7188
PutSpoonOnTableClothInScene-v0	0.8229
StackGreenCubeOnYellowCubeBakedTexInScene-v0	0.3542
Mean over all 16 run logs	0.6250

The 62.5% artifact-derived value matches the Qwen2.5 PI row in the StarVLA Model Zoo, although that document currently links the PI/GR00T rows inconsistently. This card binds the value to the logs in this repository rather than to that link.

Files

config.yaml
dataset_statistics.json
checkpoints/
├── steps_30000_pytorch_model.pt
└── steps_30000_pytorch_model_infer_*.log.run{1,2,3,4}

Loading and evaluation

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

export CKPT=$PWD/Qwen-PI-Bridge-RT-1/checkpoints/steps_30000_pytorch_model.pt
python deployment/model_server/server_policy.py \
  --ckpt_path "$CKPT" \
  --port 10093 \
  --use_bf16

Use the maintained SimplerEnv client and oxe_bridge normalization for WidowX. oxe_rt1 is the second training mixture key.

Intended use and limitations

This is the pre-2025-10-20 legacy QwenPI checkpoint for Bridge + RT-1 research, not QwenPI_v3 and not the incompatible post-restructure QwenPI layout. Its exact training code revision—and therefore whether the runtime repeat factor was 1 or 2—is not recorded. The simulation results do not establish real-robot reliability, and the policy is not safety-tuned. Observation order, state preprocessing, action normalization, and embodiment must match the packaged configuration.