LeWM world model: robosuite Panda Reach (frameskip 10)
A JEPA-style latent world model (ViT-tiny encoder, autoregressive transformer
predictor) trained on Panda Reach demonstrations with joint-velocity control.
This is epoch 18 of 20 of training run orthogonal_fs10_1h100_lowload-65566053,
the epoch with the lowest validation loss.
Files
| File | Contents |
|---|---|
weights_epoch_18.pt |
Model weights |
weights_epoch_18_object.ckpt |
Full pickled model object (torch.save(model)); load with torch.load(..., weights_only=False) with the training code on the Python path |
config.yaml |
Model architecture (Hydra config) |
Training
- Data: 1011 Panda Reach episodes × 100 steps, 224×224 RGB
agentviewframes (rendered with collision meshes visible, MuJoCo 2.3.7), 8-D joint-velocity actions - Frameskip 10 (each prediction step covers 10 env steps; action input is 10 × 8 = 80-D)
- History 3 frames, 1 predicted step, embedding dim 192
- 20 epochs, batch size 64, 1× H100
| Epoch | val loss | val pred loss |
|---|---|---|
| 17 | 0.3503 | 0.0298 |
| 18 (this checkpoint) | 0.3456 | 0.0276 |
| 19 | 0.3464 | 0.0289 |
| 20 | 0.3464 | 0.0291 |
Evaluation
Planning results reported for this run so far (50 episodes, goal 75 steps ahead, 100-step budget, success = end effector within 3 cm of the goal) were measured with the epoch 20 checkpoint: SWM+SLP 34%, SWM+CEM 12%. This epoch 18 checkpoint has not been evaluated for planning yet.
- Downloads last month
- 2