--- license: apache-2.0 library_name: openpi base_model: IliaLarchenko/behavior_submission datasets: - behavior-1k/2026-challenge-demos tags: - robotics - behavior-1k - omnigibson - pi0.5 --- # BEHAVIOR Challenge 2026, task 57 `sweeping_garage`: fine-tuned 2025 winner (expert, step 7499) This is our **submitted** policy for task 57 `sweeping_garage` (R1Pro robot, OmniGibson / Isaac Sim 5.1). It is the 2025 BEHAVIOR Challenge 1st-place model, fine-tuned on the 2026 sweeping_garage demos. ## Summary | | | |---|---| | Base checkpoint | `IliaLarchenko/behavior_submission`, `checkpoint_3` (HF revision `89545bc1b7aa7f2e687bc0032d091f132d715d4e`) | | Code | [IliaLarchenko/behavior-1k-solution](https://github.com/IliaLarchenko/behavior-1k-solution) @ `ca556f7`, plus our changes (see below) | | Data | all 200 `sweeping_garage` demos (episodes 0–199, 894,606 frames at 30 fps) from `behavior-1k/2026-challenge-demos` (LeRobot v3) | | Cameras / inputs | head + left/right wrist RGB, proprioception. Depth is not used. | | Total fine-tuning | 27,500 steps in 3 stages, batch size 32 | | Hardware | 1× RTX 5090, about 3.4 s/step | | This checkpoint | `pi_behavior_sweep_expert/sweeping_garage_expert/7499` | ## Training stages Every stage uses batch 32, a cosine LR schedule with warmup, no EMA, and the base checkpoint's own norm stats and FAST tokenizer. Each stage starts from the previous stage's final `params`. The SigLIP vision tower and the PaliGemma (Gemma-2B) backbone stay frozen in every stage. | # | Config name | Steps | Warmup | Peak LR → final LR | Trainable | Final train loss | |---|---|---|---|---|---|---| | 1 | `pi_behavior_sweep_lora` | 12,000 | 500 | 5e-5 → 5e-6 | LoRA on both Gemma towers + small heads | ~0.14 | | 2 | `pi_behavior_sweep_lora_cont` | 8,000 | 200 | 2e-5 → 2e-6 | same as stage 1 | ~0.12 | | 3 | `pi_behavior_sweep_expert` (**this model**) | 7,500 | 300 | 2e-5 → 2e-6 | stage 2 + the full Gemma-300M action expert (~531M params) | ~0.11 | "Small heads" means the task/stage embeddings, fusion, kv_transform, and action/time projections. Model settings come from the 2025 solution: action horizon 30, action dim 32, correlated flow noise (β 0.5), FAST auxiliary loss (weight 0.05), subtask loss (weight 0.1), and per-timestamp normalization. ## Changes from the 2025 code - It reads the 2026 LeRobot v3 data and uses the 2026 61-dim proprio layout (same field names as 2025). - The 2025 model has no slot for task 57, so task 57 is **trained and served in model slot 36** (`clean_a_patio`). This keeps every parameter shape identical to the checkpoint. - `base_qvel` in the state is zeroed. It was about 0 in the 2025 data, but in 2026 it holds the real velocity, which falls far outside the model's training distribution. - The FAST tokenizer was re-implemented with `tokenizers` for transformers 5 (it gives identical ids). Weights are restored with openpi's `restore_params` (newer orbax requires this). ## Exact commands The code is `src/` and `scripts/` from the training repo (also included as `policy_code/` in our submission zip). All three configs are in `src/b1k/training/config.py`. Each run was launched like this, with the openpi venv: ```bash cd b1k_winner_ft # stage 1 PYTHONPATH=src XLA_PYTHON_CLIENT_MEM_FRACTION=0.92 python scripts/train.py \ pi_behavior_sweep_lora --exp-name sweeping_garage_from_winner_ckpt3 --overwrite # stage 2 PYTHONPATH=src XLA_PYTHON_CLIENT_MEM_FRACTION=0.92 python scripts/train.py \ pi_behavior_sweep_lora_cont --exp-name sweeping_garage_cont8k --overwrite # stage 3 (this checkpoint) PYTHONPATH=src XLA_PYTHON_CLIENT_MEM_FRACTION=0.92 python scripts/train.py \ pi_behavior_sweep_expert --exp-name sweeping_garage_expert --overwrite ``` No run crashed or was resumed. The configs contain absolute pod paths for the base checkpoint (`_WINNER_CKPT`), the data root, and each stage's init `params`. Edit those paths before reusing them. ## Serving ```bash PYTHONPATH=src python scripts/serve_b1k_2026.py \ --policy-dir --config pi_behavior_sweep_expert --port 8000 ``` The server maps evaluator task id 57 to model slot 36. The 2025 per-task "eval tricks" are **off**, because slot 36's rules would force both grippers open. A **torso cap**, on by default, limits trunk joint commands to the range seen in the demos (`--no-torso-cap` disables it). Without the cap, the model often tilts the torso back and looks at the ceiling. ## Evaluation We followed the official protocol: public instances 0–9 (scene instances 301–310), 1 rollout each, the default timeout (6709 steps), `RGBDFullResWrapper`, and the torso cap on. | Run | Success | Mean q-score | Mean dirt left | |---|---|---|---| | Official run (submitted) | 1/10 (instance 310, 5552 steps) | 0.10 | 0.73 | | Repeat run, identical settings | 0/10 | 0.00 | 0.86 | The simulator is not deterministic from run to run, so treat the 1/10 as fragile. "Dirt left" is the fraction of the ~25 dirt particles still on the floor at the end. It comes from our logging hook and doesn't affect the score. ## Files - `params/`: orbax inference weights - `assets/`: norm stats used at inference time - `_CHECKPOINT_METADATA` The optimizer state (`train_state/`) is not included. ## License The base model is released under Apache-2.0. It is derived from π0.5 / PaliGemma, so the [Gemma Terms of Use](https://ai.google.dev/gemma/terms) also apply.