---
license: mit
tags:
- robotics
- imitation-learning
- diffusion-policy
- memory
- rmbench
- robotwin
---

# CAMP on RMBench

Policies for the paper **"Remember what you did: learning behavioral memory for robot manipulation"**
([CAMP](https://robo-camp.github.io/), code: https://github.com/ucsdarclab/CAMP), trained on the
[RMBench](https://github.com/robotwin-Platform/rmbench) (RoboTwin 2.0, Aloha-AgileX) benchmark from the 50 released
demonstrations per task. Evaluated with RMBench's own protocol (`demo_clean`, seeds from 100000 validated by the
scripted expert, per-task step limits, 100 episodes).

| task | success (100 episodes) | folder |
|---|---|---|
| rearrange_blocks | 100 / 100 | `rearrange_blocks/` |
| blocks_ranking_try | 100 / 100 | `blocks_ranking_try/` |
| put_back_block | 100 / 100 | `put_back_block/` |
| battery_try | 97 / 100 | `battery_try/` |

## Files

Each task folder holds only inference weights (no optimizer, scheduler or training bookkeeping):

- `policy.ckpt` — CAMP policy (Diffusion Policy conditioned on the compressed action memory), EMA weights and the
  resolved training config.
- `memory/best_model.pt` — the Stage-1 action-memory LSTM (weights + architecture args) the policy was trained with.
- `memory/normalizer.pt` — its input normaliser.

## Recipe (all tasks)

- Stage 1: memory LSTM pretrained on the 50 demos to reconstruct its past actions (DCT heads), head camera
  96x128 + 14-D joint state, hidden 128, action subsampling 4.
- Stage 2: Diffusion Policy (head camera 240x320, 14-D joint targets, `n_obs_steps=1`, 8-step action chunks) conditioned on the
  memory through a 32-D projection; memory frozen for 400 epochs, then jointly finetuned (200 epochs; put_back_block 600).
  The checkpoint reported per task is the best one over evaluated epochs.

## Usage

```bash
# inside the CAMP + RoboTwin evaluation image (see scripts/rmbench/eval in the CAMP repo)
python scripts/rmbench/eval/rmbench_eval.py eval --task rearrange_blocks --ckpt policy --episodes 100 \
    --ckpt_root <this repo>/stage2 --stage1_root <this repo>/stage1
```
where `stage2/<task>/checkpoints/policy.ckpt` and `stage1/<task>/{best_model.pt,normalizer.pt}` point at the files
of this repo (symlink or copy). The policy adapter is `scripts/rmbench/eval/policy_CAMP` and follows RMBench's
`get_model / eval / reset_model` interface.

## Citation

```bibtex
@article{wang2026rememberdidlearningbehavioral,
  title   = {Remember What You Did: Learning Behavioral Memory for Robot Manipulation},
  author  = {Wang, Kuancheng and Yeom, Hyunsoo and Cao, Yifan and Zhi, Huanyu and Shinde, Ishan and Yip, Michael C.},
  journal = {arXiv preprint arXiv:2606.21188},
  year    = {2026}
}
```