pi05_FAST_robotwin_post_training_ES

RoboTwin post-training, ES (baseline) checkpoint from Fine-Tuning VLAs with Self-Demonstrated Generative Control for Multi-Task Manipulation.

Paper · Project page · Code · RoboTwin code

• Model: π0.5 with FAST tokens
• Training data: Expert demos for the Stage-2 stack-blocks tasks (robotwin_post_training_stack_blocks_expert), fine-tuned from pi05_FAST_robotwin_mid_training.
• Step: 6000 (no optimizer state)
• Norm stats: assets/prachigarg23/robotwin_mid_training_expert/norm_stats.json

Usage

From the root of the code repo:

hf download prachigarg23/pi05_FAST_robotwin_post_training_ES --local-dir checkpoints/pi05_FAST_robotwin_post_training_ES
REPO=$(pwd); cd third_party/robotwin
CUDA_VISIBLE_DEVICES=0 python run_test_tasks.py --ckpt2 $REPO/checkpoints/pi05_FAST_robotwin_post_training_ES --cfg2 pi05_FAST_robotwin_post_training_ES \
    --log_dir logdir_test_results_H25/pi05_FAST_robotwin_post_training_ES --exec_horizon 25

See the RoboTwin README for setup and the other eval splits.

Citation

@misc{garg2026finetuningvlasselfdemonstratedgenerative,
      title={Fine-Tuning VLAs with Self-Demonstrated Generative Control for Multi-Task Manipulation},
      author={Prachi Garg and Steve Xing and Prahit Yaugand and Saurabh Gupta and Derek Hoiem},
      year={2026},
      eprint={2608.19490},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2608.19490},
}