timerewarder

TimeRewarder — MetaWorld checkpoints

Pretrained TimeRewarder checkpoints for 10 MetaWorld tasks, from the paper:

TimeRewarder: Learning Dense Reward from Passive Videos via Frame-wise Temporal Distance Yuyang Liu*, Chuan Wen*, Yihang Hu, Dinesh Jayaraman, Yang Gao†

🌐 Project Page · 📄 Paper · 💻 Code · 🤗 Demos

TimeRewarder learns a dense reward from passive (action-free) videos by predicting the frame-wise temporal distance between frames; the per-frame progress prediction is used directly as the reward for downstream RL.

Contents

One vision-only checkpoint per task (<task>_20bins.pth, 20 discrete progress bins): basketball, button-press-topdown, disassemble, door-open, drawer-open, lever-pull, plate-slide, stick-push, window-close, window-open.

huggingface-cli download CowAndSheep/timerewarder --local-dir models/ckpt

Drop the .pth files into models/ckpt/ and run the downstream RL as described in the code repository.

Citation

@article{liu2025timerewarder,
  title={TimeRewarder: Learning Dense Reward from Passive Videos via Frame-wise Temporal Distance},
  author={Liu, Yuyang and Wen, Chuan and Hu, Yihang and Jayaraman, Dinesh and Gao, Yang},
  journal={arXiv preprint arXiv:2509.26627},
  year={2025}
}