Modelpublic

StarVLA/BEHAVIOR-qwendual-state-tast1-chunck50-BEHAVIOR-rgp-seg

sourceHugging Faceupdated 24d agoView on Hugging Face

0likes47downloads

Model Card

StarVLA QwenDual for BEHAVIOR RGP Seg (70K)

This repository contains the uploaded 70,000-step checkpoint from the run named 1106_BEHAVIOR_qwendual_state_tast1_chunck50_BEHAVIOR_rgp_seg in config.yaml. The Hub repository name preserves the original tast and chunck spellings; the configured action horizon is 50.

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 / 44
Predicted action chunk	50 steps
Inference diffusion steps	4
Image resolution	224 x 224
State input	Enabled
Dataset mixture	`BEHAVIOR_rgp_seg`
Normalization key	`new_embodiment`
Uploaded checkpoint	`checkpoints/steps_70000_pytorch_model.pt`

The public configuration does not contain a task selector, camera count, or camera ordering. The repository name suggests a task-1 experiment, but that is not encoded in datasets.vla_data; this card therefore does not assert an exact task identity.

Training details

Setting	Value in `config.yaml`
Intended maximum steps	100,000
Released step	70,000
Per-device VLA batch size	16
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

Files

text

config.yaml
dataset_statistics.json
summary.jsonl
checkpoints/
└── steps_70000_pytorch_model.pt

summary.jsonl records checkpoint steps through 70K but has no evaluation values. The actual uploaded checkpoint is 70K even though the trainer target in the YAML is 100K.

Evaluation status

No success rate or episode-level evaluation artifact is published in this repository. The large counts stored in dataset_statistics.json are normalization metadata, not evaluation results, and are not used as a claimed dataset-size or performance number here.

Loading and evaluation

bash

huggingface-cli download StarVLA/BEHAVIOR-qwendual-state-tast1-chunck50-BEHAVIOR-rgp-seg \
  --local-dir BEHAVIOR-qwendual-state-tast1-chunck50-BEHAVIOR-rgp-seg

export CKPT=$PWD/BEHAVIOR-qwendual-state-tast1-chunck50-BEHAVIOR-rgp-seg/checkpoints/steps_70000_pytorch_model.pt
python deployment/model_server/server_policy.py \
  --ckpt_path "$CKPT" \
  --port 10093 \
  --use_bf16

Use the maintained BEHAVIOR integration for the simulator side. Select the packaged new_embodiment statistics and verify the 44D state and 23D action contract before inference.

Intended use and limitations

This checkpoint is for research on the configured BEHAVIOR_rgp_seg mixture. Its precise task manifest, VLM revision, camera schema, and benchmark result are not public in the Hub repository. It has not been validated for different embodiments or physical-robot deployment and is not safety-tuned.