| --- |
| license: apache-2.0 |
| library_name: openpi |
| pipeline_tag: robotics |
| tags: |
| - robotics |
| - vla |
| - pi0.5 |
| - openpi |
| - rby1 |
| - fold |
| - awbc |
| - model-soup |
| - robocup |
| --- |
| |
| # Doing Laundry — RoboCup 2026 Incheon |
|
|
| > **Team INHA-UNITED** · RB-Y1 T-shirt folding policy for the RoboCup 2026 @Home league (Incheon). |
|
|
| A π0.5 vision-language-action policy for Rainbow Robotics RB-Y1 T-shirt folding, deployed as the competition model. This is a **greedy model soup** (`M0_lr_soup`) — a weight average of a learning-rate sweep over the AWBC fold family, picked to maximize robustness on the real robot. |
|
|
| - **Method:** Model Arithmetic **greedy weight soup** of a 3-way LR sweep, all warm-started from the [v11 AWBC](https://huggingface.co/RB3159/pi05-rby1-fold-awbc) checkpoint (step 69999) and continued to step 99999. Members and soup weights: `lr1` (2.5e-6) ×2, `lr2` (5e-6) ×1, `lr4` (1e-5) ×1 — `lr1` was selected twice by greedy souping. |
| - **Training objective:** Advantage-Weighted Behavior Cloning (AWBC) following χ₀ (kai0) — per-frame `awbc_weight` from the [advantage estimator](https://huggingface.co/RB3159/rby1-fold-advantage-estimator)'s relative advantage `V(s+H) − V(s)`, thresholded at a fixed global p70 cutoff (~30% positive). See [arXiv:2602.09021](https://arxiv.org/abs/2602.09021), [OpenDriveLab/kai0](https://github.com/OpenDriveLab/kai0). |
| - **Prompt:** single task, `fold the t-shirt, Advantage: positive` (hold at inference). |
| - **Architecture:** π0.5 (`pi05`, PaliGemma `gemma_2b` + action expert `gemma_300m`, action horizon 50). |
| - **Lineage:** [pi05-rby1-fold-base](https://huggingface.co/RB3159/pi05-rby1-fold-base) (mir280) → [3stage](https://huggingface.co/RB3159/pi05-rby1-fold-3stage) → [v11 AWBC](https://huggingface.co/RB3159/pi05-rby1-fold-awbc) → LR sweep → **this soup**. |
| - **Training data:** 432 episodes — [rby1-fold-v1.1](https://huggingface.co/datasets/RB3159/rby1-fold-v1.1) master + left/right mirror + weak time-scale augmentation + DAgger interventions, with a per-frame `awbc_weight` column, img224. |
| - **Action space:** 16-dim — `[R_arm(7), R_gripper(1), L_arm(7), L_gripper(1)]`, delta on the 7-DoF arms, absolute on the grippers. |
| - **Robot:** Rainbow Robotics RB-Y1 (dual-arm). |
|
|
| ## Why a soup |
|
|
| The three LR-sweep runs share an init and a data distribution but converge to different basins. Averaging their weights (greedy soup) keeps the shared skill while smoothing over run-specific overfitting — no extra inference cost, and empirically the safest single model to deploy under competition conditions. |
|
|
| ## Files |
|
|
| `params/` (orbax inference weights) + `assets/.../norm_stats.json` + `_CHECKPOINT_METADATA`. Trained/merged with [openpi](https://github.com/Physical-Intelligence/openpi). |
|
|
| ## Usage |
|
|
| ```bash |
| python scripts/serve_policy.py --config pi05_rby1_fold_v11_1prompt_f2r03_awbc --checkpoint RB3159/Doing-Laundry-Robocup2026-Incheon |
| ``` |
|
|
| Hold the prompt to `fold the t-shirt, Advantage: positive`. |
|
|
