--- license: mit tags: - robotics - oat - robocasa - coffee_press_button --- # RoboCasa coffee — OAT TopK policy + tokenizer Handoff artifacts for mentee BoN16/32 + AWR on `coffee_press_button` (branch `robocasa`). ## Files | Path | Role | |------|------| | `policies/robocasa_coffee_press_button_topk_ep0500_sr0.600.ckpt` | BASE_CKPT (TopK @ selection seed 2000, ep 500, SR 0.60) | | `policies/robocasa_coffee_press_button_topk_lock.txt` | lock metadata | | `tokenizers/coffee_ep-1940_mse-0.003.ckpt` | frozen task tokenizer (loaded via policy) | ## Protocol See repo guide `oat/AGENT_GUIDE_ROBOCASA_COFFEE_BON16_AWR.md` in `Mirageinvo/OAT_BLT_research` (`robocasa` branch). - Report seeds: literal 5 × `10000…10004` (`-n 1`, `n_test=50`) - AWR: collect `--bon_n 16`, train `--epochs 100` - Env: `.venv_robocasa` (robosuite 1.5), `MUJOCO_GL=egl` Wave1 baseline + BoN8 eval logs / zarr: separate dataset repo (not this model repo).