Luke-Xing commited on
Commit
8d5d556
·
verified ·
1 Parent(s): 0136b5d

Upload README.md

Browse files
Files changed (1) hide show
  1. README.md +115 -0
README.md ADDED
@@ -0,0 +1,115 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ library_name: openpi
4
+ base_model: IliaLarchenko/behavior_submission
5
+ datasets:
6
+ - behavior-1k/2026-challenge-demos
7
+ tags:
8
+ - robotics
9
+ - behavior-1k
10
+ - omnigibson
11
+ - pi0.5
12
+ ---
13
+
14
+ # BEHAVIOR Challenge 2026, task 57 `sweeping_garage`: fine-tuned 2025 winner (expert, step 7499)
15
+
16
+ This is our **submitted** policy for task 57 `sweeping_garage` (R1Pro robot, OmniGibson / Isaac Sim 5.1).
17
+ It is the 2025 BEHAVIOR Challenge 1st-place model, fine-tuned on the 2026 sweeping_garage demos.
18
+
19
+ ## Summary
20
+
21
+ | | |
22
+ |---|---|
23
+ | Base checkpoint | `IliaLarchenko/behavior_submission`, `checkpoint_3` (HF revision `89545bc1b7aa7f2e687bc0032d091f132d715d4e`) |
24
+ | Code | [IliaLarchenko/behavior-1k-solution](https://github.com/IliaLarchenko/behavior-1k-solution) @ `ca556f7`, plus our changes (see below) |
25
+ | Data | all 200 `sweeping_garage` demos (episodes 0–199, 894,606 frames at 30 fps) from `behavior-1k/2026-challenge-demos` (LeRobot v3) |
26
+ | Cameras / inputs | head + left/right wrist RGB, proprioception. Depth is not used. |
27
+ | Total fine-tuning | 27,500 steps in 3 stages, batch size 32 |
28
+ | Hardware | 1× RTX 5090, about 3.4 s/step |
29
+ | This checkpoint | `pi_behavior_sweep_expert/sweeping_garage_expert/7499` |
30
+
31
+ ## Training stages
32
+
33
+ Every stage uses batch 32, a cosine LR schedule with warmup, no EMA, and the base checkpoint's own norm stats and FAST tokenizer.
34
+ Each stage starts from the previous stage's final `params`.
35
+ The SigLIP vision tower and the PaliGemma (Gemma-2B) backbone stay frozen in every stage.
36
+
37
+ | # | Config name | Steps | Warmup | Peak LR → final LR | Trainable | Final train loss |
38
+ |---|---|---|---|---|---|---|
39
+ | 1 | `pi_behavior_sweep_lora` | 12,000 | 500 | 5e-5 → 5e-6 | LoRA on both Gemma towers + small heads | ~0.14 |
40
+ | 2 | `pi_behavior_sweep_lora_cont` | 8,000 | 200 | 2e-5 → 2e-6 | same as stage 1 | ~0.12 |
41
+ | 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 |
42
+
43
+ "Small heads" means the task/stage embeddings, fusion, kv_transform, and action/time projections.
44
+ Model settings come from the 2025 solution: action horizon 30, action dim 32, correlated flow noise
45
+ (β 0.5), FAST auxiliary loss (weight 0.05), subtask loss (weight 0.1), and per-timestamp normalization.
46
+
47
+ ## Changes from the 2025 code
48
+
49
+ - It reads the 2026 LeRobot v3 data and uses the 2026 61-dim proprio layout (same field names as 2025).
50
+ - The 2025 model has no slot for task 57, so task 57 is **trained and served in model slot 36** (`clean_a_patio`).
51
+ This keeps every parameter shape identical to the checkpoint.
52
+ - `base_qvel` in the state is zeroed. It was about 0 in the 2025 data, but in 2026 it holds the real velocity,
53
+ which falls far outside the model's training distribution.
54
+ - The FAST tokenizer was re-implemented with `tokenizers` for transformers 5 (it gives identical ids).
55
+ Weights are restored with openpi's `restore_params` (newer orbax requires this).
56
+
57
+ ## Exact commands
58
+
59
+ The code is `src/` and `scripts/` from the training repo (also included as `policy_code/` in our submission zip).
60
+ All three configs are in `src/b1k/training/config.py`. Each run was launched like this, with the openpi venv:
61
+
62
+ ```bash
63
+ cd b1k_winner_ft
64
+ # stage 1
65
+ PYTHONPATH=src XLA_PYTHON_CLIENT_MEM_FRACTION=0.92 python scripts/train.py \
66
+ pi_behavior_sweep_lora --exp-name sweeping_garage_from_winner_ckpt3 --overwrite
67
+ # stage 2
68
+ PYTHONPATH=src XLA_PYTHON_CLIENT_MEM_FRACTION=0.92 python scripts/train.py \
69
+ pi_behavior_sweep_lora_cont --exp-name sweeping_garage_cont8k --overwrite
70
+ # stage 3 (this checkpoint)
71
+ PYTHONPATH=src XLA_PYTHON_CLIENT_MEM_FRACTION=0.92 python scripts/train.py \
72
+ pi_behavior_sweep_expert --exp-name sweeping_garage_expert --overwrite
73
+ ```
74
+
75
+ No run crashed or was resumed. The configs contain absolute pod paths for the base checkpoint
76
+ (`_WINNER_CKPT`), the data root, and each stage's init `params`. Edit those paths before reusing them.
77
+
78
+ ## Serving
79
+
80
+ ```bash
81
+ PYTHONPATH=src python scripts/serve_b1k_2026.py \
82
+ --policy-dir <this repo> --config pi_behavior_sweep_expert --port 8000
83
+ ```
84
+
85
+ The server maps evaluator task id 57 to model slot 36. The 2025 per-task "eval tricks" are **off**,
86
+ because slot 36's rules would force both grippers open. A **torso cap**, on by default, limits trunk
87
+ joint commands to the range seen in the demos (`--no-torso-cap` disables it). Without the cap, the model
88
+ often tilts the torso back and looks at the ceiling.
89
+
90
+ ## Evaluation
91
+
92
+ We followed the official protocol: public instances 0–9 (scene instances 301–310), 1 rollout each, the default
93
+ timeout (6709 steps), `RGBDFullResWrapper`, and the torso cap on.
94
+
95
+ | Run | Success | Mean q-score | Mean dirt left |
96
+ |---|---|---|---|
97
+ | Official run (submitted) | 1/10 (instance 310, 5552 steps) | 0.10 | 0.73 |
98
+ | Repeat run, identical settings | 0/10 | 0.00 | 0.86 |
99
+
100
+ The simulator is not deterministic from run to run, so treat the 1/10 as fragile.
101
+ "Dirt left" is the fraction of the ~25 dirt particles still on the floor at the end. It comes from our logging hook
102
+ and doesn't affect the score.
103
+
104
+ ## Files
105
+
106
+ - `params/`: orbax inference weights
107
+ - `assets/`: norm stats used at inference time
108
+ - `_CHECKPOINT_METADATA`
109
+
110
+ The optimizer state (`train_state/`) is not included.
111
+
112
+ ## License
113
+
114
+ The base model is released under Apache-2.0. It is derived from π0.5 / PaliGemma, so the
115
+ [Gemma Terms of Use](https://ai.google.dev/gemma/terms) also apply.