File size: 3,553 Bytes
a0da159 c9336dc a0da159 c9336dc 588b38a c9336dc 588b38a c9336dc b2d08c5 c9336dc fb225f3 c9336dc 744dbed c9336dc 744dbed c9336dc 744dbed fb225f3 c9336dc e22b5dd c9336dc e22b5dd c9336dc | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 | ---
license: mit
tags:
- diffusion
- robotics
- planning
- safe-control
---
# JM2D checkpoints
Pre-trained checkpoints for **Joint Model-based Model-free Diffusion for Planning with Constraints** (CoRL 2025).
[[Project page]](https://jm2d-corl25.github.io/) [[ArXiv]](https://arxiv.org/abs/2509.08775) [[Code]](https://github.com/goldenhazard/JM2D)
Wonsuhk Jung\*, Utkarsh A. Mishra\*, Nadun Ranawaka Arachchige, Yongxin Chen, Danfei Xu, Shreyas Kousik — Georgia Institute of Technology
Each folder belongs to one branch of the [code repository](https://github.com/goldenhazard/JM2D); that branch's `scripts/download_checkpoints.py` downloads
its folder and checks the md5 sums below. The toy-domain checkpoints are small and ship with the `public_toy` branch.
The hardware branch (`public_hardware`) is a guide for running JM2D on your own robot and has no checkpoint.
## Files
| Folder | Branch | Files | Size |
|---|---|---|---|
| `pointmaze/` | `public_robomimic` | `maze2d_large_ep{1250,1500,1750,1900,2000}.pth`: robomimic diffusion policies on D4RL `maze2d-large-dense-v1` (the 5 checkpoints of paper Fig. 4) | 5 × 524 MB |
| `d3il/data/avoiding/` | `public_d3il` | `env_*.pkl` (96 files) + `MD5SUMS`: the D3IL Avoiding demonstrations, needed for training and for evaluation (data normalizer) | 2.6 MB |
| `d3il/avoiding-d3il/H8_K20_Dmodels.GaussianDiffusion/<seed>/` | `public_d3il` | `state_best.pt` + config pickles: Diffuser-style planners on D3IL-Avoiding, seeds 0–4 (paper Tables 2 and 3) | 5 × 32 MB |
| `pointmaze/dataset/` | `public_robomimic` | `maze2d_large_dense_v1.hdf5`: the D4RL `maze2d-large-dense-v1` dataset converted to robomimic format, used to train the PointMaze policies | 303 MB |
| File | md5 |
|---|---|
| `pointmaze/maze2d_large_ep1250.pth` | `98614093985811a0dbcbb4a66b3ccfc4` |
| `pointmaze/maze2d_large_ep1500.pth` | `000b0f287307bd98289cc9b28479566c` |
| `pointmaze/maze2d_large_ep1750.pth` | `a7fd3fe32e1b35723dde1c566c61f1b9` |
| `pointmaze/maze2d_large_ep1900.pth` | `a756af86fe2070c2e77358ca380ccc44` |
| `pointmaze/maze2d_large_ep2000.pth` | `7f76546cb300ec7577c475d1f3b40f97` |
| `d3il/.../0/state_best.pt` | `8446a72feebe0e9b1839bc5cb48f448e` |
| `d3il/.../1/state_best.pt` | `c09dd9a85fcf435201d8e61835d053f0` |
| `d3il/.../2/state_best.pt` | `4c55388894fdf2ebeb22206678125ac3` |
| `d3il/.../3/state_best.pt` | `7d219c1b2df29557297bde3ab561b784` |
| `d3il/.../4/state_best.pt` | `7238d20b674872da4a8d88e8469a5a3b` |
| `pointmaze/dataset/maze2d_large_dense_v1.hdf5` | `aa4be41d7444c0bdc8fcf48707e11c84` |
## Download
From a checkout of the matching branch:
```bash
python scripts/download_checkpoints.py
```
or directly:
```bash
hf download RoboWonsuhk/JM2D --include "pointmaze/*" --local-dir checkpoints_hf
```
The PointMaze dataset is a converted copy of the D4RL dataset (Fu et al., 2020, https://github.com/Farama-Foundation/D4RL), redistributed under its CC BY 4.0 license.
The D3IL Avoiding demonstrations are an unmodified copy from D3IL (Jia et al., 2024, https://github.com/ALRhub/d3il, MIT License).
The files are PyTorch pickles; load them only with the code of the matching branch.
## Citation
```bibtex
@inproceedings{jung2025joint,
title={Joint Model-based Model-free Diffusion for Planning with Constraints},
author={Wonsuhk Jung and Utkarsh Aashu Mishra and Nadun Ranawaka Arachchige and Yongxin Chen and Danfei Xu and Shreyas Kousik},
booktitle={9th Annual Conference on Robot Learning},
year={2025},
url={https://openreview.net/forum?id=E9t1ekt6W9}
}
```
|