Datasets:
AdaJEPA datasets
Training and test data for AdaJEPA: An Adaptive Latent World Model (Ying Wang, Oumayma Bounou, Yann LeCun, Mengye Ren). Paper: arXiv:2606.32026. Website: agenticlearning.ai/adajepa.
Each environment has its own top-level folder: pushobj/ and pointmaze/.
Abstract
Latent world models enable planning from high-dimensional observations by predicting future states in a compact latent space. However, these models are typically kept frozen at test time: when their predictions become inaccurate, planning can fail, especially under test-time distribution shift. To address this, we propose AdaJEPA, an adaptive latent world model that performs test-time adaptation within the closed loop of model predictive control (MPC). After training, AdaJEPA plans and executes the first action chunk, uses the observed next-state transition as a self-supervised adaptation signal, and replans with the updated model. This closed-loop update continuously recalibrates the world model without additional expert demonstrations. Across a range of goal-reaching tasks, AdaJEPA substantially improves planning success with as few as one gradient step per MPC replanning step.
Download
hf download yingw/adajepa --repo-type dataset --include "pushobj/PushObj-TLZ+-4k/*" --local-dir data
tar -xf data/pushobj/PushObj-TLZ+-4k/obses.tar -C data/pushobj/PushObj-TLZ+-4k
PushObj training folders with 16,000 episodes ship their videos as one uncompressed
obses.tar, since the Hub allows at most 10,000 files per directory. PushObj test folders
ship obses/ directly. The PointMaze folder ships its observation tensors as
obses.tar.zst; extract it with tar --zstd -xf.
PushObj
PushObj is a PushT variant with seven block shapes: T, L, Z, +, square, I and
small_tee. Trajectories are regenerated per shape with a bias toward contact, so the
pusher touches the block in most episodes. Each shape has 16,000 training trajectories.
Test trajectories are held-out episodes with at least one contact.
| Subset | Contents | Size | Paper use |
|---|---|---|---|
PushObj-TLZ+-4k |
4,000 trajectories for each of T, L, Z, + |
1.8 GB | shape-shift experiments |
PushObj-T-16k |
16,000 trajectories of T |
1.8 GB | visual-shift experiments |
PushObj-all-16k |
16,000 trajectories per shape, one folder per shape | 12 GB | data-scaling experiments |
PushObj-test |
test trajectories and planning targets | 2.0 GB | all PushObj evaluations |
PushObj-T-16k is identical to PushObj-all-16k/T. PushObj-TLZ+-4k is a separate
generation and only its T block matches the first 4,000 episodes of PushObj-all-16k/T.
Folder layout
Every folder, including each shape of PushObj-all-16k and each val_* test set, has
the same files. E is the number of episodes in the folder.
| File | Shape and type | Meaning |
|---|---|---|
obses/episode_NNN.mp4 |
224×224 MPEG-4, 10 fps | one frame per step, one file per episode |
abs_actions.pth |
float32 [E, 246, 2] |
absolute pusher target position |
rel_actions.pth |
float32 [E, 246, 2] |
per-step action, the model input |
states.pth |
float32 [E, 246, 5] |
pusher x, y; block x, y; block angle |
velocities.pth |
float32 [E, 246, 2] |
pusher velocity from the simulator |
seq_lengths.pkl |
list of E ints |
true episode length, 49 to 246 |
shapes.pkl |
list of E strings |
block shape of each episode |
Episodes are padded to 246 steps. Steps past seq_lengths[i] are padding, and the video
of episode i has exactly seq_lengths[i] frames.
Test sets
| Folder | Episodes | Condition |
|---|---|---|
val |
100 | 25 each of T, L, Z, +, for monitoring training loss |
val_T, val_L, val_Z, val_+ |
2,706 each | shapes seen by PushObj-TLZ+-4k |
val_square, val_I, val_small_tee |
2,706 each | held-out shapes |
The visual-shift conditions (blur, salt-and-pepper noise, dark lighting, red block,
anchor or agent) are not stored. The release code renders them from val_T at
evaluation time through its ood_corruption and env_kwargs_override options.
Each val_* folder also holds plan_targets.pkl, the 1,000 planning tasks the paper
evaluates on (seed 42, traj_len 25). Each entry of segments is one task:
| Key | Value |
|---|---|
ep_idx |
source episode in this folder |
offset |
step at which the task starts |
states |
float32 [26, 5], start state through the goal state 25 steps later |
actions |
float32 [25, 2], expert actions over that window |
shape |
block shape |
The start image is frame offset of the episode and the goal image is frame offset + 25.
PointMaze-Medium
pointmaze/PointMaze-Medium-4k holds 4,000 trajectories of 100 steps in the medium maze
(maze2d-medium-v0), used for the dynamics-shift experiments. It is one pool: the
training code splits it 90/10 into 3,600 training and 400 validation trajectories with a
fixed seed. Test episodes are not drawn from it. Planning starts and goals are sampled
from the 26 free cells of the maze, accepting a pair only when its distance exceeds
3 cell units, with uniformly random initial velocities.
| File | Shape and type | Meaning |
|---|---|---|
obses.tar.zst |
4,000 files obses/episode_NNNN.pth, uint8 [100, 224, 224, 3] |
one frame tensor per episode, 57 GB extracted |
actions.pth |
float64 [4000, 100, 2] |
forces along x and y |
states.pth |
float64 [4000, 100, 4] |
position and velocity of the point |
seq_lengths.pth |
int64 [4000] |
all 100 |
metadata.pth |
dict | n_trajs, traj_len, maze_type, env_id |
Citation
@misc{wang2026adajepa,
title = {AdaJEPA: An Adaptive Latent World Model},
author = {Ying Wang and Oumayma Bounou and Yann LeCun and Mengye Ren},
year = {2026},
eprint = {2606.32026},
archivePrefix = {arXiv},
primaryClass = {cs.LG},
url = {https://arxiv.org/abs/2606.32026}
}
- Downloads last month
- 332