MimicGen Panda/UR5e 4-task balanced checkpoints
Full DART/OpenPI checkpoints trained separately for the Panda and UR5e embodiments on the same task-balanced four-task setup.
Training data per embodiment
| Task | Episodes |
|---|---|
| Square | 950 |
| Threading | 950 |
| Stack | 475 |
| Stack Three | 475 |
| Total | 2,850 |
Both embodiment runs use normalization statistics computed from the combined
Panda and UR5e source data. The uploaded folders include params, assets,
train_state, and Orbax checkpoint metadata, so they can be used for inference,
fine-tuning, or exact optimizer-state resume.
Repository layout
panda/
9999/
12999/
14999/
ur5e/
9999/
12999/
14999/
The directory number is the zero-based saved training step. For example,
14999 is the checkpoint saved after 15,000 optimization steps.