robomimic Diffusion Policy checkpoints
Diffusion Policy (UNet) checkpoints trained with robomimic 0.5.0 on low-dimensional observations
from the proficient-human (ph) v1.5 datasets. Both runs used 2000 epochs and seed 1.
| Task | Environment | Checkpoint | Success rate (50 episodes) |
|---|---|---|---|
| Lift | Lift |
lift/model_epoch_2000.pth |
0.98 |
| Can | PickPlaceCan |
can/model_epoch_2000.pth |
0.98 |
The robot is a Panda with an OSC_POSE delta controller at 20 Hz. Each task folder includes the
training config.json. Rollout videos and a summary are on the
diffusion branch of q2rl.
To load a checkpoint, use robomimic.utils.file_utils.policy_from_checkpoint. Each file is about
1 GB, is a pickled PyTorch checkpoint, and should only be loaded if you trust its source.