File size: 3,109 Bytes
e7f35dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
57a4466
e7f35dc
691e02c
e7f35dc
691e02c
 
 
 
e7f35dc
 
 
 
 
691e02c
 
 
e7f35dc
691e02c
 
 
e7f35dc
 
691e02c
 
 
e7f35dc
 
 
691e02c
 
 
 
 
e7f35dc
 
 
 
 
 
 
 
 
 
 
 
 
49d569b
 
e7f35dc
691e02c
e7f35dc
691e02c
e7f35dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
---
license: cc-by-4.0
pretty_name: PointMaze Medium (Temporal Straightening)
task_categories:
  - robotics
  - reinforcement-learning
tags:
  - world-model
  - planning
  - point-maze
  - maze2d
  - d4rl
  - temporal-straightening
size_categories:
  - 1K<n<10K
---

# PointMaze Medium

The PointMaze-Medium dataset from
[Temporal Straightening for Latent Planning](https://arxiv.org/abs/2603.12231)
(Wang et al., 2026). It has 4,000 trajectories of 100 steps in the D4RL
`maze2d-medium-v0` maze, rendered at 224×224. It extends the umaze set in
[DINO-WM](https://huggingface.co/datasets/gaoyuezhou/dino_wm) and uses the same
file layout.

## Layout

```
point_maze_medium/
├── obses/episode_000.pth … episode_3999.pth  # uint8 [100, 224, 224, 3] per episode
├── states.pth        # float64 [4000, 100, 4]: x, y, vx, vy
├── actions.pth       # float64 [4000, 100, 2]: in [-1, 1]
├── seq_lengths.pth   # int64 [4000], all 100
├── metadata.pth      # generation settings
├── examples/         # mp4 clips of the first ten episodes
└── MANIFEST.sha256   # checksums: sha256sum -c MANIFEST.sha256
```

Frame `t` and `states[:, t]` are both taken after step `t`; `t = 0` is the
reset. There is no stored split. The paper splits by trajectory at load time,
3,600 train and 400 validation.

## Generation

The point-maze environment from the
[temporal-straightening code](https://github.com/agentic-learning-ai-lab/temporal-straightening),
which builds on DINO-WM (d4rl 1.1, mujoco-py 2.1.2.14, MuJoCo 2.1.0, gym 0.23.1).
Actions are uniform in [-1, 1]² with numpy seed 0. Episode `i` starts from
`env.sample_random_init_goal_states(100 * i)`.

## Loading

```python
import torch
from huggingface_hub import snapshot_download

root = snapshot_download("agentic-learning-ai-lab/point_maze_medium", repo_type="dataset")
states = torch.load(f"{root}/states.pth")              # [4000, 100, 4]
actions = torch.load(f"{root}/actions.pth")            # [4000, 100, 2]
frames = torch.load(f"{root}/obses/episode_000.pth")  # uint8 [100, 224, 224, 3]
```

To use it with the temporal-straightening code, place or symlink the downloaded
directory at `$DATASET_DIR/point_maze_medium`.

## License

CC BY 4.0. Please cite the paper below if you use the data.

## Citation

```bibtex
@article{wang2026temporal_straightening,
  title={Temporal Straightening for Latent Planning},
  author={Wang, Ying and Bounou, Oumayma and Zhou, Gaoyue and Balestriero, Randall and Rudner, Tim GJ and LeCun, Yann and Ren, Mengye},
  journal={arXiv preprint arXiv:2603.12231},
  year={2026}
}

@article{zhou2024dinowm,
  title={DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning},
  author={Zhou, Gaoyue and Pan, Hengkai and LeCun, Yann and Pinto, Lerrel},
  journal={arXiv preprint arXiv:2411.04983},
  year={2024}
}

@article{fu2020d4rl,
  title={D4RL: Datasets for Deep Data-Driven Reinforcement Learning},
  author={Fu, Justin and Kumar, Aviral and Nachum, Ofir and Tucker, George and Levine, Sergey},
  journal={arXiv preprint arXiv:2004.07219},
  year={2020}
}
```