---
license: mit
tags:
- robotics
- oat
- robocasa
---

# RoboCasa last-two policies + tokenizers

Datasets + Wave1 logs: [`hackhackhack66666/rc-last-two`](https://huggingface.co/datasets/hackhackhack66666/rc-last-two)

Guides (branch `robocasa`):
- [SINK BoN16/AWR](https://github.com/Mirageinvo/OAT_BLT_research/blob/robocasa/oat/AGENT_GUIDE_ROBOCASA_SINK_BON16_AWR.md)
- [MICROWAVE BoN16/AWR](https://github.com/Mirageinvo/OAT_BLT_research/blob/robocasa/oat/AGENT_GUIDE_ROBOCASA_MICROWAVE_BON16_AWR.md)

## Inventory

| File | Role |
|------|------|
| `policies/robocasa_turn_off_sink_faucet_topk_ep0500_sr0.580.ckpt` | locked TopK policy (sink) |
| `policies/robocasa_turn_off_sink_faucet_topk_lock.txt` | lock metadata |
| `policies/robocasa_turn_off_microwave_topk_ep0500_sr0.620.ckpt` | locked TopK policy (microwave) |
| `policies/robocasa_turn_off_microwave_topk_lock.txt` | lock metadata |
| `tokenizers/sink_ep-3080_mse-0.002.ckpt` | frozen OAT tokenizer |
| `tokenizers/microwave_ep-2720_mse-0.002.ckpt` | frozen OAT tokenizer |

## Mentee: enough for AWR16?

**Yes**, together with `rc-last-two` (zarr + **full** Wave1 baselines for both tasks):

1. Copy policies → `my_models/`
2. Place tokenizers at the paths in the guides (`output/20260720/.../checkpoints/`)
3. Untar zarrs → `data/robocasa/`
4. Rsync `wave1_*` → `output/eval/matched_s10000/robocasa/<task>/`
5. Follow guide: BoN16/32 eval → collect `--bon_n 16` → train `--epochs 100`

Wave1 BoN8 for microwave may still be running — **not required** to start AWR16.