Σ-LeWM checkpoints
Inference checkpoints for Σ-LeWM: Learning World Models via S-Information for Robot Control.
| Directory | Task | Training seed | Macro-action dimension |
|---|---|---|---|
pusht/ |
Push-T | 3072 | 10 |
tworoom/ |
TwoRoom | 2 | 10 |
reacher/ |
DMControl Reacher | 3072 | 10 |
cube/ |
OGBench Cube | 3072 | 25 |
scene/ |
OGBench Scene | 1 | 25 |
Each directory contains weights.pt and its matching config.json. These are
single-checkpoint inference exports, not the complete multi-seed experiment bundle.
Training-only backward, inverse and regularisation modules are omitted. No datasets,
evaluation records or experimental scores are included. File checksums are in SHA256SUMS.
Usage
Install the linked code repository and its dependencies, and prepare the task datasets as described there. Download the weights:
export STABLEWM_HOME=/path/to/stablewm
hf download zuxinrui/sigma-lewm --local-dir "$STABLEWM_HOME/checkpoints/sigma-lewm"
slewm eval --task pusht --checkpoint sigma-lewm/pusht/weights.pt --horizon 5 --cem 30 --execute
Replace pusht in both arguments with tworoom, reacher, cube or scene.
The configuration targets the slewm package from the linked repository.
Weights are plain state dictionaries and can be read with
torch.load(path, map_location="cpu", weights_only=True).