File size: 5,406 Bytes
b32d742 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 | diff --git a/src/openpi/configs/robots/b1k.py b/src/openpi/configs/robots/b1k.py
index 0fa8121..f6c9006 100644
--- a/src/openpi/configs/robots/b1k.py
+++ b/src/openpi/configs/robots/b1k.py
@@ -4,24 +4,26 @@ from .base_config import ObservationConfig, StateActionConfig, RobotConfig, regi
# R1Pro Robot Configuration
# Dual-arm mobile manipulator with base, torso, and multiple camera views
R1Pro = RobotConfig(
- name="robot",
+ # BEHAVIOR-1K v3.9.2 eval names the robot "robot_r1" (see OmniGibson/omnigibson/eval/r1pro.yaml),
+ # so the flattened observation keys the eval client streams are prefixed "robot_r1::".
+ name="robot_r1",
robot_type="R1Pro",
observations={
"image_0": ObservationConfig(
name="head",
- obs_key="robot::robot:zed_link:Camera:0::rgb",
+ obs_key="robot_r1::robot_r1:zed_link:Camera:0::rgb",
dataset_key="observation.rgb.zed_link_camera_0",
resolution=[240, 240]
),
"image_1": ObservationConfig(
name="left_wrist",
- obs_key="robot::robot:left_realsense_link:Camera:0::rgb",
+ obs_key="robot_r1::robot_r1:left_realsense_link:Camera:0::rgb",
dataset_key="observation.rgb.left_realsense_link_camera_0",
resolution=[240, 240]
),
"image_2": ObservationConfig(
name="right_wrist",
- obs_key="robot::robot:right_realsense_link:Camera:0::rgb",
+ obs_key="robot_r1::robot_r1:right_realsense_link:Camera:0::rgb",
dataset_key="observation.rgb.right_realsense_link_camera_0",
resolution=[240, 240]
),
diff --git a/src/openpi/configs/tasks/b1k.py b/src/openpi/configs/tasks/b1k.py
index 193c4e7..e492b58 100644
--- a/src/openpi/configs/tasks/b1k.py
+++ b/src/openpi/configs/tasks/b1k.py
@@ -3,6 +3,13 @@ from . import TASK_REGISTRY
# i3l task prompts
TASKS = {
"turning_on_radio": "Turn on the radio receiver that's on the table in the living room.",
+ # Task 22 in the 2026 BEHAVIOR Challenge task list. The policy was fine-tuned with
+ # prompt_from_task=True, so the training prompt was the raw LeRobot task string
+ # "putting_shoes_on_rack" (the PaliGemma tokenizer lowercases it and turns "_" into " ").
+ "putting_shoes_on_rack": "putting shoes on rack",
+ # Task 38 in the 2026 BEHAVIOR Challenge. Same convention: prompt_from_task=True during
+ # fine-tuning, so the training prompt is the raw LeRobot task string.
+ "spraying_for_bugs": "spraying for bugs",
}
# Register in global registry
diff --git a/src/openpi/training/config.py b/src/openpi/training/config.py
index ec6cd0f..7200a0b 100644
--- a/src/openpi/training/config.py
+++ b/src/openpi/training/config.py
@@ -772,6 +772,61 @@ _CONFIGS = [
assets_base_dir="./outputs/assets",
checkpoint_base_dir="./outputs/checkpoints",
),
+ # Task 38 (spraying_for_bugs) -- full fine-tune, mirror of pi05_b1k. Needs multi-GPU (FSDP)
+ # or a large single GPU; use the _lora variant below to fit on one ~46 GB card.
+ TrainConfig(
+ name="pi05_b1k_spraying_for_bugs",
+ model=pi0_config.Pi0Config(action_horizon=32, pi05=True),
+ data=LeRobotB1KDataConfig(
+ repo_id="spraying_for_bugs",
+ base_config=DataConfig(
+ data_cls=_lerobot_compat.LeRobotDataset,
+ dataset_root="/opt/b1k/datasets/spraying_for_bugs",
+ prompt_from_task=True,
+ dataset_kwargs={"tolerance_s": 5e-4},
+ ),
+ robot_config_name="b1k/R1Pro",
+ ),
+ weight_loader=weight_loaders.CheckpointWeightLoader("gs://openpi-assets/checkpoints/pi05_base/params"),
+ save_interval=10_000,
+ num_train_steps=50_000,
+ assets_base_dir="./outputs/assets",
+ checkpoint_base_dir="./outputs/checkpoints",
+ ),
+ # Task 38 (spraying_for_bugs) -- LoRA fine-tune. Fits on a single L40S (~46 GB), even alongside
+ # the eval policy server. Lower ceiling than the full fine-tune above.
+ TrainConfig(
+ name="pi05_b1k_spraying_for_bugs_lora",
+ model=pi0_config.Pi0Config(
+ action_horizon=32,
+ pi05=True,
+ paligemma_variant="gemma_2b_lora",
+ action_expert_variant="gemma_300m_lora",
+ ),
+ data=LeRobotB1KDataConfig(
+ repo_id="spraying_for_bugs",
+ base_config=DataConfig(
+ data_cls=_lerobot_compat.LeRobotDataset,
+ dataset_root="/opt/b1k/datasets/spraying_for_bugs",
+ prompt_from_task=True,
+ dataset_kwargs={"tolerance_s": 5e-4},
+ ),
+ robot_config_name="b1k/R1Pro",
+ ),
+ weight_loader=weight_loaders.CheckpointWeightLoader("gs://openpi-assets/checkpoints/pi05_base/params"),
+ freeze_filter=pi0_config.Pi0Config(
+ action_horizon=32,
+ pi05=True,
+ paligemma_variant="gemma_2b_lora",
+ action_expert_variant="gemma_300m_lora",
+ ).get_freeze_filter(),
+ ema_decay=None,
+ save_interval=10_000,
+ num_train_steps=50_000,
+ batch_size=16,
+ assets_base_dir="./outputs/assets",
+ checkpoint_base_dir="./outputs/checkpoints",
+ ),
#
# Fine-tuning Libero configs.
#
|