Dongkkka/Task_000759_Move_Vegetable_Whiteroom_TurboVLA-bs8-step5000

SG2 Task000759 TurboVLA — step 5000

TurboVLA checkpoint trained on Task_000759_Move_Vegetable_Whiteroom_Leardboard_MCAP_lerobot_v30 for the SG2 robot.

Model contract

• Checkpoint: optimizer step 5,000, batch size 8
• Observations: left-head, left-wrist, and right-wrist RGB cameras plus 22-D state
• Actions: 22-D absolute joint/base action
• Action horizon: 50
• Recommended inference weights: steps_5000_ema_pytorch_model.pt
• Runtime source revision: a3365e8f1787d30bafd99d2806edbf4dc82b5370

config.yaml and dataset_statistics.json must remain in the same directory as the downloaded checkpoint. The StarVLA loader discovers them from the checkpoint's parent directory.

Download

hf download Dongkkka/Task_000759_Move_Vegetable_Whiteroom_TurboVLA-bs8-step5000 \
  --local-dir ./turbovla_task000759_step5000

Inference on the prepared container2 checkout

cd /NHNHOME/WORKSPACE/26motir001_D/intern_team/Leaderboard_intern/turbovla_official

bash scripts/robotwin/run_policy_server.sh \
  /absolute/path/turbovla_task000759_step5000/steps_5000_ema_pytorch_model.pt \
  0 5694

On another machine, install the matching TurboVLA/StarVLA runtime and override the BERT and DINOv3 locations when starting server_policy.py. The packaged config records the original container2 paths. A full checkpoint load can skip GroundingDINO initialization by passing:

--cfg-option framework.text.bert_path=/path/to/bert-base-uncased \
--cfg-option framework.vision.model_path=/path/to/dinov3-vitl \
--cfg-option framework.initialization.load_pretrained=false

Files and integrity

• steps_5000_ema_pytorch_model.pt — EMA inference checkpoint, SHA-256 93f5fefc871470e9d6c979e12cb4dbcbf68be2ebd15041e9806ef501bedf5127
• steps_5000_pytorch_model.pt — raw training checkpoint, SHA-256 db2e409231021c7ff2c76337dc823990370aa5466fa955a48126006d51441d48
• config.yaml — loader configuration
• config.full.yaml — complete resolved training configuration
• dataset_statistics.json — state/action normalization statistics

This artifact has not been claimed as real-robot validated; it is the saved training checkpoint.