StarVLA/1115_BEHAVIOR_rgp_dual_QwenDual_task6_40k

StarVLA QwenDual for BEHAVIOR RGP Dual, task 6 (40K)

This repository contains the uploaded 40,000-step checkpoint from the 1115_BEHAVIOR_rgp_dual_QwenDual_task6 run. The packaged config.yaml describes a StarVLA QwenDual policy trained on task 6 of the BEHAVIOR_rgp_dual mixture. The configured training target is 100,000 steps, but the only weight file published here is the 40,000-step checkpoint.

Model details

Item	Published configuration
Framework	StarVLA QwenDual
VLM	Local training snapshot named Qwen3-VL-4B-Instruct; revision not recorded
Auxiliary visual encoder	dinov2_vits14
Action model	16-layer DiT-B flow head: 768 latent width, 12 heads (64 dimensions/head); state/action decoder MLP width 1,024
Action / state dimension	23 / 21
Predicted action chunk	50 steps
Image resolution	224 x 224
State input	Enabled
Dataset mixture / task	BEHAVIOR_rgp_dual, task_id: 6
Normalization key	R1Pro
Uploaded checkpoint	checkpoints/steps_40000_pytorch_model.pt

The public configuration does not declare the number or ordering of camera views. Preserve the training-time observation adapter and the packaged normalization statistics instead of constructing a flat input manually. The current StarVLA BEHAVIOR documentation defines the 23D action as base, torso, two 7D arms, and two grippers.

Training details

Setting	Value in config.yaml
Intended maximum steps	100,000
Released step	40,000
Per-device VLA batch size	32
Gradient accumulation	1
Warm-up steps	5,000
Base / interface / action LR	4e-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
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 lists saved steps through 40K. It is a checkpoint index, not an evaluation report.

Files

config.yaml
dataset_statistics.json
summary.jsonl
checkpoints/
└── steps_40000_pytorch_model.pt

Keep config.yaml and dataset_statistics.json beside the checkpoint. The StarVLA server uses them to reconstruct the model and apply the R1Pro normalization contract.

Evaluation status

No BEHAVIOR success rate, challenge score, episode manifest, or per-task evaluation log is published in this repository. Consequently this card does not attach a benchmark number to the 40K artifact. The public BEHAVIOR example is still marked under construction and should be treated as an integration guide rather than evidence that this checkpoint reproduces a challenge entry.

Loading and evaluation

This is a StarVLA framework checkpoint, not a Transformers from_pretrained() directory.

huggingface-cli download StarVLA/1115_BEHAVIOR_rgp_dual_QwenDual_task6_40k \
  --local-dir 1115_BEHAVIOR_rgp_dual_QwenDual_task6_40k

export CKPT=$PWD/1115_BEHAVIOR_rgp_dual_QwenDual_task6_40k/checkpoints/steps_40000_pytorch_model.pt
python deployment/model_server/server_policy.py \
  --ckpt_path "$CKPT" \
  --port 10093 \
  --use_bf16

Use the maintained BEHAVIOR evaluation guide for the simulator client. Verify state dimension, camera order, action order, and normalization key at the server handshake before running episodes.

Intended use and limitations

This checkpoint is intended for BEHAVIOR-1K simulation research with the R1Pro embodiment and the task/data scope recorded above. It has no published evaluation result, no recorded upstream VLM revision, and no demonstrated transfer to other tasks, embodiments, or physical robots. BEHAVIOR evaluation requires a Vulkan-capable simulator GPU; the project documentation warns that A100/H100-class GPUs without RT cores are not suitable for this simulator. Do not deploy the policy on hardware without an independent safety layer.