File size: 5,455 Bytes
8d5d556
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
---
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 <this repo> --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.