---
library_name: pytorch
tags:
  - robotics
  - imitation-learning
  - diffusion-policy
  - libero
---

# StreamingWAM LIBERO checkpoint

This repository contains the released StreamingWAM checkpoint trained on the
four standard LIBERO suites and evaluated on LIBERO-Plus. The checkpoint is
intended for use with the accompanying
[StreamingWAM code](https://github.com/renshaojie233/StreamingWAM).

## File

| File | Size | SHA-256 |
| --- | ---: | --- |
| `streamingwam_libero.pt` | 12,042,077,420 bytes | `35499c8b2ac7bc879c988c9af4f9e9ff9052caccd22d582b7fadd90de185d496` |

The PyTorch checkpoint contains the model weights (`mot` and
`proprio_encoder`) and minimal loading metadata (`step` and `torch_dtype`). It
does not contain optimizer state, scheduler state, random-number-generator
state, training logs, or experiment-tracking data.

## Usage

Download `streamingwam_libero.pt`, install the StreamingWAM codebase, and pass
the local checkpoint path to the evaluation command:

```bash
python experiments/libero/eval_libero_task_list_multi_k.py \
  task=streamingwam_libero_plus \
  ckpt=/path/to/streamingwam_libero.pt \
  EVALUATION.dataset_stats_path=assets/libero_dataset_stats.json \
  +EVALUATION.task_list_file=assets/libero_plus_full_10030.txt \
  +EVALUATION.sdp_k_values='[4]'
```

See the code repository for the complete environment, data layout, and
evaluation options.

## Scope and terms

This checkpoint is released for research on robot learning and policy
evaluation. The source code is MIT-licensed. External datasets, pretrained
components, benchmark assets, and model weights retain their respective
licenses and terms.