notmuch2/eVTA0_datasets

eVTA₀ Datasets

This repository hosts the training and evaluation data of the paper "Demonstration-Free Success-Probability Reward Learning for Generalist Robot Policies":

• Rollout collection (data/libero/): 2000 mixed-quality LIBERO policy rollouts with soft failure targets (V-JEPA2 + DBSCAN + sigmoid augmentation). 1600 trajectories are used for training eVTA₀ (890 success / 710 failure); the remaining 400 form the eval split, used for evaluation.
• Trajectory split (train/libero/libero_trajectory_split.csv): per-rollout train / eval assignment with the success flag.
• Evaluation selections (eval/libero/): the demonstration trajectories used for VOC / VROC (demonstrations/progress_trajectory_ids.csv, 400 trajectories, 100 per suite) and the eval rollout manifest for MSE / Kendall tau-a (policy_rollouts/policy400_terminal_manifest.csv, 400 rollouts; file paths relative to data/libero).

The VOC / VROC demonstrations are public LeRobot datasets, downloaded separately from the Hugging Face Hub: lerobot/libero_spatial_image, lerobot/libero_object_image, lerobot/libero_goal_image, lerobot/libero_10_image (episode_index in the selection CSV is the per-suite episode number).

MetaWorld: TODO.

Data format

The rollout collection is one HDF5 file per task with one traj_* group per rollout (obs_main_images, obs_wrist_images, rewards, dones, actions); rewards[-1] is the terminal target — 1.0 for successes and a soft value in (0, 1) for failures (see data/libero/augmentation_report.json and ../eVTA0/evta0/evta0/augment.py).

Code: https://github.com/duowuyms/eVTA0

Paper: https://arxiv.org/abs/2609.33653

Author: Duo Wu

Citation

If you find these datasets useful, please cite our paper:

@article{wu2026evta0,
  title={Demonstration-Free Success-Probability Reward Learning for Generalist Robot Policies},
  author={Wu, Duo and Wang, Haifeng and Lu, Rongwei and Wang, Jinghe and Xiong, Tianyi and Wang, Zhimin and Yu, Chao and Ma, Shuai and Wang, Zhi},
  journal={arXiv preprint arXiv:2609.33653},
  year={2026}
}