---
license: apache-2.0
tags:
  - robotics
  - reinforcement-learning
  - humanoid
---

# FADA Oracle PPO Checkpoints

Oracle PPO expert checkpoints for **FADA: Few-Shot Domain Adaptation via Dynamics Alignment for
Humanoid Control** (arXiv:2606.28476), used as the `--expert-checkpoint` input to the DAgger
distillation step of the FADA pipeline.

- Code: https://github.com/lecar-lab/FADA
- Paper: https://arxiv.org/abs/2606.28476
- Project page: https://lecar-lab.github.io/FADA-humanoid/

## Contents

- `t1/`: Booster T1 (23-DoF) oracle PPO training run, every intermediate checkpoint saved every
  100 iterations (`model_00000.pt` ... `model_24999.pt`, 251 files).
- `g1/`: Unitree G1 (29-DoF) oracle PPO training run, every intermediate checkpoint saved every
  100 iterations (`model_00000.pt` ... `model_24999.pt`, 251 files).

The full intermediate history is included, not just the final checkpoint: FADA's DAgger training
reward-samples 20 checkpoints from across the oracle run for its weak-policy ("suboptimal-data")
source, so `--expert-checkpoint` must point into a directory holding every intermediate
`model_*.pt`. See the code repository's README for the exact usage and how to opt out of that
source (`--suboptimal-data-ratio 0`), which only requires the single final checkpoint.

## Security note

A `.pt` checkpoint is executable content. The FADA codebase loads these with torch's restricted
`weights_only=True` reader, which raises rather than executing an untrusted pickle stream.

## Citation

```bibtex
@article{xie2026fada,
  title   = {FADA: Few-Shot Domain Adaptation via Dynamics Alignment for Humanoid Control},
  author  = {Xie, Angchen and Sobanbabu, Nikhil and Shikhare, Ishayu and Wang, Alan
             and Simchowitz, Max and Shi, Guanya},
  journal = {arXiv preprint arXiv:2606.28476},
  year    = {2026},
  doi     = {10.48550/arXiv.2606.28476}
}
```