MINT-SJTU/Evo-DexBench_StarVLA_checkpoints

StarVLA × Evo-DexBench — 15-task joint checkpoints

Six policies, one per model and embodiment. Each is trained jointly on all 15 single-arm or all 15 dual-arm Evo-DexBench tasks: 624 demonstrations, 50 from each of 12 Core tasks and 8 from each of 3 Gen tasks, sampled with equal task weights. Policies use 30D/60D physical actions, 31D/62D state and three RGB cameras, and run at 30 Hz.

Case folder	Framework	Training steps	Action horizon / executed steps
qwenoft-single, qwenoft-dual	QwenOFT (official StarVLA framework)	600k	50 / 10
wan-single, wan-dual	EvoWanOFT (model extension)	300k	8 / 8
cosmos-single, cosmos-dual	EvoCosmosGR00T (model extension)	300k	8 / 8

Per-group checkpoints trained on four or five tasks are in Travor278/Evo-DexBench-StarVLA-100k.

Download and run

Download one case plus the model-serving package:

hf download MINT-SJTU/Evo-DexBench_StarVLA_checkpoints \
  --include "qwenoft-single/*" "model_extension/*" "SHA256.json" --local-dir ./models

Each case folder contains final_model/pytorch_model.pt, config.yaml and dataset_statistics.json. SHA256.json lists the size and SHA-256 of every file.

Initialize the public StarVLA source at 4507931a625c844404c4a76128fb116536d8ca7c (or a later revision that contains it) and set STARVLA_ROOT to that checkout. Install model_extension in the model environment with python -m pip install -e ./models/model_extension --no-deps.

QwenOFT requires the original Qwen3-VL base and processor:

hf download Qwen/Qwen3-VL-4B-Instruct \
  --revision ebb281ec70b05090aa6165b016eac8ec08e71b17 \
  --local-dir ./models/base_models/Qwen3-VL-4B-Instruct

Start the QwenOFT model server:

python -m evodex_wm4a.serve --framework QwenOFT \
  --ckpt-path ./models/qwenoft-single/final_model/pytorch_model.pt --port 10101 \
  --config-override "framework.qwenvl.base_vlm=$PWD/models/base_models/Qwen3-VL-4B-Instruct"

Use a Qwen3-VL-capable model environment (Transformers 4.57.0). Wan and Cosmos also need the matching base_configs/* folder: base_configs/wan or base_configs/Cosmos-Predict2-2B-Video2World. Their server is started the same way with --framework EvoWanOFT or --framework EvoCosmosGR00T. The 15-task mixture shares its robot type with the per-group mixtures, so serve it with --config-override datasets.vla_data.data_mix=evodex_single_h50_q99 (QwenOFT) or evodex_single_h8_q99 (Wan/Cosmos). Use dual in place of single for the dual-arm cases. The saved dataset_statistics.json is used unchanged. The base_configs/ paths in the Wan and Cosmos config.yaml files are relative, so start their server from the download directory or override framework.evodex_wan.base_config_dir (or framework.evodex_cosmos.base_config_dir and base_config_alias). Training data is not distributed with the checkpoints.

Evaluate with the public Evo-DexBench benchmark and its StarVLA policy plugin.

Training

All runs use global batch 4, seed 42, bf16, 5,000 warmup steps, cosine decay to 5e-7 and gradient clipping of 1.0 on NVIDIA A100 80 GB GPUs.

Model	Learning rate, base / action head	GPUs x per-device batch	Recipe
QwenOFT	1e-5 / 1e-4	2 x 2	recipes/{single,dual}15_qwenoft_h50_q99.yaml with trainer.max_train_steps=600000
Wan-OFT	1e-4 / 1e-4	1 x 4	recipes/{single,dual}15_wan_h8_mean_q99.yaml
Cosmos-GR00T	1e-4 / 1e-4	1 x 4	recipes/{single,dual}15_cosmos_h8_mean_q99.yaml

The QwenOFT backbone is frozen. The Wan and Cosmos backbones are frozen and their action heads are trained. config.yaml in each case folder is the training configuration with machine-specific paths replaced. SHA256.json records each replaced value.

Evaluation

Each checkpoint is evaluated on all 15 tasks of its embodiment with the Evo-DexBench standard protocol: 100 episodes per task in five conditions of 20 (Nominal, Camera, Lighting, Physical, Spatial), layouts 34/33/33, seeds 1001–1005, four parallel environments, 30 Hz control, 120 Hz simulation and the step limit in the task YAML. SR is the success rate and Stage is the mean normalized stage score; values are fractions.

Case	SR, 15 tasks	Stage, 15 tasks	SR, 12 Core tasks	SR, 3 Gen tasks
qwenoft-single	0.157	0.257	0.173	0.090
wan-single	0.121	0.198	0.146	0.020
cosmos-single	0.129	0.241	0.151	0.043
qwenoft-dual	0.018	0.109	0.023	0.000
wan-dual	0.005	0.099	0.006	0.003
cosmos-dual	0.037	0.168	0.044	0.010

M01–M04 were re-evaluated with the updated M01/M03 button physics and stage scoring of the public benchmark. Mean SR / Stage over these four tasks: qwenoft-single 0.277 / 0.412, wan-single 0.280 / 0.428, cosmos-single 0.230 / 0.380.

evaluation_results.json lists every task with SR, Stage, per-condition SR, step limit and the SHA-256 of the checkpoint and episode log.

Sources

Upstream terms apply per model family: StarVLA, Qwen3-VL, Wan policy, Cosmos policy, Wan base, Cosmos base.