fix: stabilize human-inloop dataset schema

这个提交包含在:
Elvinyk
2026-03-19 10:35:31 +08:00
父节点 0b720b1e0b
当前提交 57af3361ce
共有 3 个文件被更改,包括 207 次插入 和 16 次删除

查看文件

@ -0,0 +1,107 @@
#!/usr/bin/env python
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Patch legacy human-in-loop datasets so they can merge with newer rollout datasets."""
import logging
from dataclasses import dataclass
from pathlib import Path
import numpy as np
from lerobot.configs import parser
from lerobot.datasets.dataset_tools import add_features
from lerobot.datasets.lerobot_dataset import LeRobotDataset
from lerobot.scripts.lerobot_record import _ensure_human_inloop_compatible_features
from lerobot.utils.constants import ACTION
from lerobot.utils.utils import init_logging
MISSING_HIL_FEATURES = (
"complementary_info.policy_action",
"complementary_info.is_intervention",
"complementary_info.state",
)
@dataclass
class PatchHilDatasetSchemaConfig:
repo_id: str
root: str | None = None
output_repo_id: str | None = None
output_dir: str | None = None
def _build_missing_hil_features(dataset: LeRobotDataset) -> dict[str, tuple[np.ndarray, dict]]:
action_feature = dataset.features[ACTION]
action_names = action_feature["names"]
if action_names is None:
action_names = [f"action_{idx}" for idx in range(action_feature["shape"][0])]
else:
action_names = list(action_names)
feature_defs: dict[str, dict] = {}
_ensure_human_inloop_compatible_features(feature_defs, action_feature_names=action_names)
num_frames = dataset.num_frames
num_action_dims = len(action_names)
values_by_feature = {
"complementary_info.policy_action": np.zeros((num_frames, num_action_dims), dtype=np.float32),
"complementary_info.is_intervention": np.zeros((num_frames, 1), dtype=np.float32),
"complementary_info.state": np.zeros((num_frames, 1), dtype=np.float32),
}
missing_features = {}
for feature_name in MISSING_HIL_FEATURES:
if feature_name not in dataset.features:
missing_features[feature_name] = (values_by_feature[feature_name], feature_defs[feature_name])
return missing_features
@parser.wrap()
def patch_hil_dataset_schema(cfg: PatchHilDatasetSchemaConfig) -> LeRobotDataset:
init_logging()
dataset = LeRobotDataset(cfg.repo_id, root=cfg.root)
missing_features = _build_missing_hil_features(dataset)
if not missing_features:
logging.info("Dataset already has a merge-compatible human-in-loop schema. Nothing to patch.")
return dataset
output_repo_id = cfg.output_repo_id or f"{cfg.repo_id}_hil_schema_patched"
output_dir = Path(cfg.output_dir) if cfg.output_dir is not None else None
if output_dir is None and cfg.root is not None:
src_root = Path(cfg.root)
output_dir = src_root.parent / f"{src_root.name}_hil_schema_patched"
logging.info("Adding missing human-in-loop schema features: %s", sorted(missing_features))
patched_dataset = add_features(
dataset=dataset,
features=missing_features,
output_dir=output_dir,
repo_id=output_repo_id,
)
logging.info("Patched dataset saved to %s", patched_dataset.root)
return patched_dataset
def main():
patch_hil_dataset_schema()
if __name__ == "__main__":
main()

查看文件

@ -295,6 +295,29 @@ class RecordConfig:
return ["policy"]
def _ensure_human_inloop_compatible_features(
dataset_features: dict[str, dict],
*,
action_feature_names: list[str],
) -> None:
# Keep human-in-loop datasets schema-stable across teleop-only and policy-assisted phases so they can merge.
dataset_features["complementary_info.policy_action"] = {
"dtype": "float32",
"shape": (len(action_feature_names),),
"names": action_feature_names,
}
dataset_features["complementary_info.is_intervention"] = {
"dtype": "float32",
"shape": (1,),
"names": ["is_intervention"],
}
dataset_features["complementary_info.state"] = {
"dtype": "float32",
"shape": (1,),
"names": ["state"],
}
@parser.wrap()
def record(cfg: RecordConfig) -> LeRobotDataset:
init_logging()
@ -330,24 +353,10 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
use_videos=cfg.dataset.video,
),
)
if cfg.intervention_state_machine_enabled and cfg.policy is not None and cfg.teleop is not None:
if cfg.teleop is not None and cfg.enable_collector_policy_id:
action_names = dataset_features[ACTION]["names"]
action_names = list(robot.action_features) if action_names is None else list(action_names)
dataset_features["complementary_info.policy_action"] = {
"dtype": "float32",
"shape": (len(action_names),),
"names": action_names,
}
dataset_features["complementary_info.is_intervention"] = {
"dtype": "float32",
"shape": (1,),
"names": ["is_intervention"],
}
dataset_features["complementary_info.state"] = {
"dtype": "float32",
"shape": (1,),
"names": ["state"],
}
_ensure_human_inloop_compatible_features(dataset_features, action_feature_names=action_names)
if cfg.enable_collector_policy_id:
dataset_features["complementary_info.collector_policy_id"] = {
"dtype": "string",

查看文件

@ -22,6 +22,7 @@ import numpy as np
import pytest
import torch
from lerobot.datasets.dataset_tools import merge_datasets, remove_feature
from lerobot.datasets.lerobot_dataset import LeRobotDataset
from lerobot.scripts.lerobot_calibrate import CalibrateConfig, calibrate
from lerobot.scripts.lerobot_human_inloop_record import (
@ -31,6 +32,7 @@ from lerobot.scripts.lerobot_human_inloop_record import (
_slow_reset_all_arms_to_pose,
human_inloop_record,
)
from lerobot.scripts.lerobot_patch_hil_dataset_schema import PatchHilDatasetSchemaConfig, patch_hil_dataset_schema
from lerobot.scripts.lerobot_record import (
ACPInferenceConfig,
DatasetRecordConfig,
@ -163,13 +165,86 @@ def test_human_inloop_record_works_without_policy_and_saves_annotations(tmp_path
assert cfg.intervention_state_machine_enabled is False
assert cfg.collector_policy_id_policy == "human"
assert "complementary_info.collector_policy_id" in dataset.features
assert "complementary_info.policy_action" in dataset.features
assert "complementary_info.is_intervention" in dataset.features
assert "complementary_info.state" in dataset.features
reloaded = LeRobotDataset(DUMMY_REPO_ID, root=root)
assert reloaded[0]["complementary_info.collector_policy_id"] == "human"
torch.testing.assert_close(
reloaded[0]["complementary_info.policy_action"],
torch.zeros_like(reloaded[0]["action"]),
)
assert float(reloaded[0]["complementary_info.is_intervention"]) == 0.0
assert float(reloaded[0]["complementary_info.state"]) == 0.0
assert "episode_success" in reloaded.meta.episodes.column_names
assert reloaded.meta.episodes[0]["episode_success"] == "failure"
def test_patch_hil_dataset_schema_restores_legacy_dataset_mergeability(tmp_path):
robot_cfg = MockRobotConfig()
teleop_cfg = MockTeleopConfig()
current_root = tmp_path / "hil_current"
legacy_root = tmp_path / "hil_legacy"
patched_root = tmp_path / "hil_patched"
merged_root = tmp_path / "hil_merged"
dataset_cfg = DatasetRecordConfig(
repo_id=DUMMY_REPO_ID,
single_task="Dummy task",
root=current_root,
num_episodes=1,
episode_time_s=0.1,
reset_time_s=0,
push_to_hub=False,
)
cfg = RecordConfig(
robot=robot_cfg,
dataset=dataset_cfg,
teleop=teleop_cfg,
play_sounds=False,
)
current_dataset = human_inloop_record(cfg)
legacy_dataset = remove_feature(
current_dataset,
feature_names=[
"complementary_info.policy_action",
"complementary_info.is_intervention",
"complementary_info.state",
],
output_dir=legacy_root,
repo_id="dummy/repo_legacy",
)
assert "complementary_info.policy_action" not in legacy_dataset.features
patched_dataset = patch_hil_dataset_schema(
PatchHilDatasetSchemaConfig(
repo_id="dummy/repo_legacy",
root=str(legacy_root),
output_repo_id="dummy/repo_patched",
output_dir=str(patched_root),
)
)
assert "complementary_info.policy_action" in patched_dataset.features
assert "complementary_info.is_intervention" in patched_dataset.features
assert "complementary_info.state" in patched_dataset.features
patched_reloaded = LeRobotDataset("dummy/repo_patched", root=patched_root)
torch.testing.assert_close(
patched_reloaded[0]["complementary_info.policy_action"],
torch.zeros_like(patched_reloaded[0]["action"]),
)
assert float(patched_reloaded[0]["complementary_info.is_intervention"]) == 0.0
assert float(patched_reloaded[0]["complementary_info.state"]) == 0.0
merged_dataset = merge_datasets(
datasets=[current_dataset, patched_dataset],
output_repo_id="dummy/repo_merged",
output_dir=merged_root,
)
assert merged_dataset.meta.total_episodes == current_dataset.meta.total_episodes + patched_dataset.meta.total_episodes
def test_record_loop_sets_leader_manual_control_during_reset():
class MockTeleopWithManualControl(MockTeleop):
def __init__(self, config):