# KinDER Checkpoints

Final trained checkpoints for reproducing all KinDER evaluation results.

## Directory Structure

```
checkpoints/kinder/
  {env}/
    planner/state_{step}.pt   # Diffusion planner checkpoint
    invdyn/state_{step}.pt    # Inverse dynamics model checkpoint
```

## Checkpoint Inventory

| Environment | Planner | InvDyn | Notes |
|---|---|---|---|
| Motion2D-p0 | state_505000.pt (1.4G) | state_800000.pt (8.4M) | |
| BaseMotion3D | state_1995000.pt (2.0G) | state_400000.pt (8.5M) | |
| DynObstruction2D-o1 | state_1995000.pt (2.0G) | state_400000.pt (34M) | |
| StickButton2D-b1 | state_855000.pt (1.4G) | state_400000.pt (8.6M) | |
| DynPushPullHook2D-o5 | state_1995000.pt (2.0G) | state_400000.pt (37M) | |
| Transport3D-o2 | state_580000.pt (2.7G) | state_400000.pt (34M) | Demo tracking bypasses planner |
| SweepIntoDrawer3D | — | state_400000.pt (51M) | Planner not needed (demo tracking) |
| TidyBot3D-cupboard | state_615000.pt (2.7G) | state_400000.pt (33M) | Demo tracking bypasses planner |

**Total: 15 files, ~15GB** (stored via Git LFS)

## Usage

These checkpoints are loaded automatically by `eval_kinder.py` when you specify the correct `--planner_epoch` and `--invdyn_epoch` flags, with `--log_dir` pointing to a directory containing the expected `{env}/planner/{run_name}/` and `{env}/invdyn/{run_name}/` structure.

To use these checkpoints directly:
```bash
# Symlink into logs/ structure expected by eval
ln -s $(pwd)/checkpoints/kinder logs/kinder_ckpts

# Or copy to the expected path
cp -r checkpoints/kinder/BaseMotion3D/ logs/kinder-BaseMotion3D-v0/
```

See `docs/kinder_eval_guide.md` for full reproduction commands.