---
license: mit
tags:
  - robotics
  - reinforcement-learning
  - world-model
  - robomimic
  - jax
---

# Pretrained state world models

Anonymous supplementary checkpoints for the accompanying double-blind
submission. These state-space dynamics models are used for tree search on
Robomimic low-dimensional tasks.

Each `.pkl` contains the Flax parameter tree for a frozen dynamics module. It
does not contain policy, critic, optimizer, dataset, or evaluation outputs.
Use the checkpoint whose seed matches the training run.

| Directory | Task and dataset | Pretraining steps |
| --- | --- | ---: |
| `lift_fixed10/` | Lift, PH fixed-10 subset | 100,000 |
| `can_mh/` | Can, MH | 100,000 |
| `square/` | Square, PH | 100,000 |
| `tool_hang_inlinewm200k/` | Tool Hang, PH | 200,000 |

Download all files with:

```bash
hf download qlearningwithworldmodels/anonymous-state-world-models \
  --include "*.pkl" \
  --local-dir checkpoints
```

Then pass a matching local file to `WM_CKPT` or
`--wm_dynamics_ckpt_path`. Verify downloads against `SHA256SUMS` before
loading. Python pickle files can execute code during deserialization; only
load files obtained from this repository with matching hashes.

Author and archival links are intentionally omitted during double-blind
review and will be added after the review period.