---
license: mit
tags:
  - robotics
  - robomimic
  - oat
  - imitation-learning
---

# RoboMimic OAT — Table P tokenizers & policies

Frozen **OAT tokenizers**, **base policies**, and **AWR (BoN-distill N=8)** checkpoints for RoboMimic Lift / Can / Square matched Table P.

## Layout

```
tokenizers/{lift,can,square}/ep-*_mse-*.ckpt
policies/{lift,can,square}/ep-*_sr-*.ckpt
awr/awr_s10000_*.ckpt
```

## Which ckpt

| Suite | Tokenizer | Base policy | AWR |
|-------|-----------|-------------|-----|
| **Can** | `tokenizers/can/ep-0520_mse-0.005.ckpt` | `policies/can/ep-1700_sr-0.940.ckpt` | `awr/awr_s10000_can.ckpt` |
| **Lift** | `tokenizers/lift/ep-1970_mse-0.006.ckpt` | `policies/lift/ep-0900_sr-0.930.ckpt` | `awr/awr_s10000_lift.ckpt` |
| **Square** | `tokenizers/square/ep-0690_mse-0.004.ckpt` | `policies/square/ep-0700_sr-0.420.ckpt` | `awr/awr_s10000_square.ckpt` |

BoN has no separate checkpoint — run inference BoN (`--bon_free 8 --bon_signal vote`) on the **base** policy.

## Companion data

Zarr: [`hackhackhack66666/robomimic_zarr`](https://huggingface.co/datasets/hackhackhack66666/robomimic_zarr)

Eval also needs RoboMimic mh image HDF5 for env metadata.