chore(ci): make quality checks pass across repo

- fix ruff violations in dataset, recording, and value modules

- fix typos findings and markdown/workflow lint issues

- update pages workflow checkout to disable credential persistence

- apply pre-commit auto-formatting and EOF normalization

- include requested repo housekeeping updates (.gitignore and README)
这个提交包含在:
Elvin
2026-03-04 11:02:06 +08:00
父节点 908eb4bbbc
当前提交 3a87382ba4
共有 27 个文件被更改,包括 152 次插入 和 120 次删除

查看文件

@ -29,6 +29,8 @@ jobs:
steps:
- name: Checkout
uses: actions/checkout@v4
with:
persist-credentials: false
- name: Setup Pages
uses: actions/configure-pages@v5

1
.gitignore vendored
查看文件

@ -177,3 +177,4 @@ outputs/
*.urdf
*.xml
*.part
AGENTS.md

查看文件

@ -22,7 +22,7 @@ If using `conda activate /llm_jzm/cache/conda_env/lerobot`:
- Treat it as shared machine mode.
- Hugging Face cache is not in the default location.
- Set Hugging Face token to the shared token for runs on this machine.
- Set Wandb token to hte shared token for runs on this machine
- Set Wandb token to the shared token for runs on this machine
Use:

查看文件

@ -1,7 +1,7 @@
<h1 align="center">Evo-RL</h1>
<p align="center">
<a href="https://elvin-yk.github.io/Evo-RL/"><img alt="project website" src="https://img.shields.io/badge/Project-Website-0ea5e9"/></a>
<a href="https://MINT-SJTU.github.io/Evo-RL/"><img alt="project website" src="https://img.shields.io/badge/Project-Website-0ea5e9"/></a>
<a href="https://github.com/huggingface/lerobot"><img alt="lerobot version" src="https://img.shields.io/badge/LeRobot-0.4.4-f59e0b"/></a>
<a href="https://evorl.example.com/wechat-post"><img alt="wechat post" src="https://img.shields.io/badge/WeChat-Official%20Post-07c160"/></a>
<a href="#community-channels"><img alt="wechat group join us" src="https://img.shields.io/badge/WeChat%20Group-Join%20Us-a855f7?logo=wechat&logoColor=white"/></a>
@ -84,6 +84,7 @@ For SO101 setup, please follow the [official SO101 tutorial](https://wiki.seeeds
#### Device path recommendation
Recommended path strategy:
- **Robot serial:** use `/dev/serial/by-id/` (stable across reboots).
- **Cameras:** prefer `/dev/v4l/by-id/`; if IDs are not unique, use `/dev/v4l/by-path/`.
- In examples below: robot ports use `by-id`, camera paths use `by-path`.
@ -213,6 +214,7 @@ lerobot-human-inloop-record \
Recommendation: use **`fourcc: "MJPG"`** for OpenCV and **`warmup_s`** for RealSense. In this example `front` uses RealSense, but you can switch it to OpenCV with the same structure.
Hotkeys:
- `i`: toggle intervention mode (policy <-> teleop takeover)
- `s`: mark success and end current episode
- `f`: mark failure and end current episode
@ -230,7 +232,7 @@ This prints: dataset meta, totals, episode-length stats/histogram, success/inter
### 4) Value Function Training
Train the value function on the current dataset. Current default: [Pi*0.6](https://www.pi.website/blog/pistar06) (`--value.type=pistar06`).
Train the value function on the current dataset. Current default: [Pi\*0.6](https://www.pi.website/blog/pistar06) (`--value.type=pistar06`).
Single-GPU template:
@ -260,6 +262,7 @@ CUDA_VISIBLE_DEVICES=<GPU_ID_LIST> accelerate launch \
```
To plug in a different value function, minimal path in this repo:
- Add `src/lerobot/values/<your_value>/configuration_<your_value>.py` with `@PreTrainedConfig.register_subclass("<your_value>")`.
- Add `src/lerobot/values/<your_value>/modeling_<your_value>.py` with `<YourValue>Policy(PreTrainedPolicy)` (implement at least `forward`, `predict_value`, and `build_training_raw_batch_hook` for `lerobot-value-train`).
- Add `src/lerobot/values/<your_value>/processor_<your_value>.py` with `make_<your_value>_pre_post_processors(...)`.
@ -268,6 +271,7 @@ To plug in a different value function, minimal path in this repo:
### 5) Value Inference
Infer value signals and write value/advantage/indicator back to the dataset:
- `value`: estimated return-to-go of the current frame.
- `advantage`: relative improvement signal (higher means better-than-baseline trajectory quality).
- `indicator`: binarized training tag derived from advantage.
@ -302,10 +306,12 @@ CUDA_VISIBLE_DEVICES=<GPU_ID_LIST> accelerate launch \
```
Parameter notes:
- `--acp.n_step`: n-step advantage horizon.
- `--acp.positive_ratio`: positive label ratio after advantage binarization (e.g., `0.3` = top 30% per task).
Expected new columns:
- `complementary_info.value_<TAG>`
- `complementary_info.advantage_<TAG>`
- `complementary_info.acp_indicator_<TAG>`
@ -313,6 +319,7 @@ Expected new columns:
These columns are written back to the original dataset specified by `--dataset.repo_id`.
Visualization (optional):
- Enable with `--viz.enable=true`.
- Output videos are saved to `outputs/value_infer/<RUN_NAME>/value/viz/`.
- Each video overlays a value curve and per-frame text (`advantage`, `acp_indicator`) on top of camera frames.
@ -323,6 +330,7 @@ Visualization (optional):
- `--viz.vcodec=<CODEC>` and `--viz.frame_storage_mode=memory|disk`
Value overlay examples (SO101, smooth=11):
- GIF assets are stored under `website/assets/gifs/`.
### 6) Policy Training
@ -354,6 +362,7 @@ lerobot-train \
`--acp.indicator_dropout_prob` controls tag drop rate in task text; `0.3` helps learn both tagged and untagged conditions.
Important checks:
- `--acp.indicator_field` must exist in the dataset and be **binary (`0/1`)**.
Multi-GPU template:
@ -396,6 +405,7 @@ lerobot-human-inloop-record \
```
Dataset continuation options:
- **Append in place:** keep `--resume=true` and continue recording into the same dataset.
- **Merge multiple rounds:** use the official dataset editor to merge separate datasets.
@ -407,6 +417,7 @@ lerobot-edit-dataset \
```
Additional data attributes vs default `lerobot-record` behavior:
- `complementary_info.policy_action`: policy output action at each step.
- `complementary_info.is_intervention`: whether current step is in intervention.
- `complementary_info.state`: intervention state-machine state.

查看文件

@ -41,7 +41,9 @@ class ValueTargetsConfig:
if self.c_fail_coef < 0:
raise ValueError("'targets.c_fail_coef' must be non-negative.")
if not self.target_field.startswith("observation."):
raise ValueError("'targets.target_field' must start with 'observation.' to survive processor conversion.")
raise ValueError(
"'targets.target_field' must start with 'observation.' to survive processor conversion."
)
@dataclass
@ -113,7 +115,7 @@ class ValueTrainPipelineConfig(HubMixin):
self.targets.validate()
if hasattr(self.value, "target_key"):
setattr(self.value, "target_key", self.targets.target_field)
self.value.target_key = self.targets.target_field
if not self.job_name:
self.job_name = f"{self.value.type}_value"
@ -136,7 +138,9 @@ class ValueTrainPipelineConfig(HubMixin):
self.scheduler = self.value.get_scheduler_preset()
if self.value.push_to_hub and not self.value.repo_id:
raise ValueError("'value.repo_id' argument missing. Please specify it to push the model to the hub.")
raise ValueError(
"'value.repo_id' argument missing. Please specify it to push the model to the hub."
)
@classmethod
def __get_path_fields__(cls) -> list[str]:

查看文件

@ -836,7 +836,11 @@ def _copy_and_reindex_episodes_metadata(
for key in schema.names:
if key in {"episode_index", "tasks", "length"}:
continue
if key.startswith("stats/") or key.startswith("__index_level_") or key.startswith("meta/episodes/"):
if (
key.startswith("stats/")
or key.startswith("__index_level_")
or key.startswith("meta/episodes/")
):
continue
if key not in seen_preserved_keys:
seen_preserved_keys.add(key)

查看文件

@ -729,15 +729,14 @@ class LeRobotDataset(torch.utils.data.Dataset):
# Check if cached dataset contains all requested episodes
if not self._check_cached_episodes_sufficient():
raise FileNotFoundError("Cached dataset doesn't contain all requested episodes")
except (AssertionError, FileNotFoundError, NotADirectoryError):
except (AssertionError, FileNotFoundError, NotADirectoryError) as err:
if is_valid_version(self.revision):
self.revision = get_safe_version(self.repo_id, self.revision)
self.download(download_videos)
self.hf_dataset = self.load_hf_dataset()
if not self._check_cached_episodes_sufficient():
missing_videos_hint = (
" Missing video files are required by this dataset. "
"Retry with `download_videos=True`."
" Missing video files are required by this dataset. Retry with `download_videos=True`."
if len(self.meta.video_keys) > 0 and not download_videos
else ""
)
@ -745,7 +744,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
"Cached dataset is incomplete after download. "
"Some requested episodes or files are missing."
f"{missing_videos_hint}"
)
) from err
# Create mapping from absolute indices to relative indices when only a subset of the episodes are loaded
# Build a mapping: absolute_index -> relative_index_in_filtered_dataset

查看文件

@ -60,10 +60,7 @@ def _from_meta_stats(dataset: Any, indicator_field: str) -> ACPIndicatorStats |
count_raw = _stats_entry_to_scalar(field_stats, "count")
total_count = int(round(count_raw)) if count_raw is not None else -1
if total_count >= 0:
positive_count = int(round(ratio * total_count))
else:
positive_count = -1
positive_count = int(round(ratio * total_count)) if total_count >= 0 else -1
return ACPIndicatorStats(
indicator_field=indicator_field,

查看文件

@ -136,7 +136,9 @@ def build_report(dataset_root: Path) -> dict[str, Any]:
episodes_df = episodes_ds.to_pandas()
actual_episode_count = int(len(episodes_df))
episode_success_labels = episodes_df["episode_success"].tolist() if "episode_success" in episodes_df else []
episode_success_labels = (
episodes_df["episode_success"].tolist() if "episode_success" in episodes_df else []
)
normalized_labels = [_normalize_label(v) for v in episode_success_labels]
success_count = sum(1 for v in normalized_labels if v == "success")
@ -153,7 +155,11 @@ def build_report(dataset_root: Path) -> dict[str, Any]:
intervention_frames = 0
intervention_episode_count = 0
if intervention_col in data_schema_names and episode_index_col in data_schema_names and actual_frame_count > 0:
if (
intervention_col in data_schema_names
and episode_index_col in data_schema_names
and actual_frame_count > 0
):
table = data_dataset.to_table(columns=[episode_index_col, intervention_col])
episode_indices = table[episode_index_col].to_pylist()
intervention_values = table[intervention_col].to_pylist()
@ -190,10 +196,14 @@ def build_report(dataset_root: Path) -> dict[str, Any]:
fps = float(info.get("fps", 0) or 0)
episode_lengths_frames = (
episodes_df["length"].astype(int).tolist() if "length" in episodes_df and actual_episode_count > 0 else []
episodes_df["length"].astype(int).tolist()
if "length" in episodes_df and actual_episode_count > 0
else []
)
episode_lengths_seconds = [frames / fps for frames in episode_lengths_frames] if fps > 0 else []
length_mean = float(sum(episode_lengths_seconds) / len(episode_lengths_seconds)) if episode_lengths_seconds else 0.0
length_mean = (
float(sum(episode_lengths_seconds) / len(episode_lengths_seconds)) if episode_lengths_seconds else 0.0
)
length_min = float(min(episode_lengths_seconds)) if episode_lengths_seconds else 0.0
length_max = float(max(episode_lengths_seconds)) if episode_lengths_seconds else 0.0
length_histogram = _build_episode_length_histogram(episode_lengths_seconds, bins=20)

查看文件

@ -22,16 +22,16 @@ Phase A/B goals:
- Keyboard-toggled intervention lets teleop temporarily take over execution.
"""
import logging
import json
import logging
import time
from pathlib import Path
from typing import Any
from lerobot.configs import parser
from lerobot.scripts.lerobot_record import RecordConfig, record
from lerobot.utils.import_utils import register_third_party_plugins
from lerobot.utils.constants import HF_LEROBOT_HOME
from lerobot.utils.import_utils import register_third_party_plugins
from lerobot.utils.recording_annotations import EPISODE_FAILURE, infer_collector_policy_version
@ -64,7 +64,9 @@ def _load_failure_reset_pose(pose_path: Path) -> dict[str, float]:
payload = json.load(f)
joint_pos_raw = payload["joint_pos"] if isinstance(payload, dict) and "joint_pos" in payload else payload
if not isinstance(joint_pos_raw, dict):
raise ValueError(f"Invalid failure reset pose payload in {pose_path}: expected dict, got {type(joint_pos_raw)}")
raise ValueError(
f"Invalid failure reset pose payload in {pose_path}: expected dict, got {type(joint_pos_raw)}"
)
joint_pos = {str(key): float(value) for key, value in joint_pos_raw.items() if str(key).endswith(".pos")}
if not joint_pos:
raise ValueError(f"Invalid failure reset pose payload in {pose_path}: no '.pos' joints found.")

查看文件

@ -97,6 +97,13 @@ from lerobot.robots import ( # noqa: F401
so_follower,
unitree_g1 as unitree_g1_robot,
)
from lerobot.scripts.recording_hil import (
ACPInferenceConfig,
PolicySyncDualArmExecutor,
_capture_policy_runtime_state, # noqa: F401
_predict_policy_action_with_acp_inference, # noqa: F401
)
from lerobot.scripts.recording_loop import record_loop
from lerobot.teleoperators import ( # noqa: F401
TeleoperatorConfig,
bi_openarm_leader,
@ -110,13 +117,6 @@ from lerobot.teleoperators import ( # noqa: F401
so_leader,
unitree_g1,
)
from lerobot.scripts.recording_hil import (
ACPInferenceConfig,
PolicySyncDualArmExecutor,
_capture_policy_runtime_state, # noqa: F401
_predict_policy_action_with_acp_inference, # noqa: F401
)
from lerobot.scripts.recording_loop import record_loop
from lerobot.utils.constants import ACTION
from lerobot.utils.control_utils import (
init_keyboard_listener,
@ -288,6 +288,7 @@ class RecordConfig:
"""This enables the parser to load config from the policy using `--policy.path=local/dir`"""
return ["policy"]
@parser.wrap()
def record(cfg: RecordConfig) -> LeRobotDataset:
init_logging()
@ -325,10 +326,7 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
)
if cfg.intervention_state_machine_enabled and cfg.policy is not None and cfg.teleop is not None:
action_names = dataset_features[ACTION]["names"]
if action_names is None:
action_names = list(robot.action_features)
else:
action_names = list(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),),
@ -524,9 +522,7 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
continue
extra_episode_metadata = (
{"episode_success": episode_success}
if cfg.enable_episode_outcome_labeling
else None
{"episode_success": episode_success} if cfg.enable_episode_outcome_labeling else None
)
dataset.save_episode(extra_episode_metadata=extra_episode_metadata)
recorded_episodes += 1

查看文件

@ -35,6 +35,9 @@ from lerobot.configs.value import ValueInferencePipelineConfig
from lerobot.datasets.lerobot_dataset import LeRobotDataset
from lerobot.datasets.utils import load_info, write_info
from lerobot.policies.factory import make_policy, make_pre_post_processors
from lerobot.scripts.value_infer_viz import (
_export_overlay_videos,
)
from lerobot.utils.constants import (
CHECKPOINTS_DIR,
LAST_CHECKPOINT_LINK,
@ -44,11 +47,6 @@ from lerobot.utils.import_utils import register_third_party_plugins
from lerobot.utils.random_utils import set_seed
from lerobot.utils.recording_annotations import EPISODE_SUCCESS, resolve_episode_success_label
from lerobot.utils.utils import init_logging, inside_slurm
from lerobot.scripts.value_infer_viz import (
_export_overlay_videos,
_get_episode_value_bounds,
_get_video_encode_options,
)
from lerobot.values.pistar06.configuration_pistar06 import Pistar06Config
from lerobot.values.pistar06.modeling_pistar06 import (
EpisodeTargetInfo,
@ -310,8 +308,8 @@ def _binarize_advantages(
for i in range(advantages.shape[0]):
task_idx = int(task_indices[i])
thr = thresholds[task_idx]
indicators[i] = 1 if float(advantages[i]) >= thr else 0
threshold = thresholds[task_idx]
indicators[i] = 1 if float(advantages[i]) >= threshold else 0
if force_intervention_positive:
intervention_mask = interventions.astype(np.float32) > 0.5

查看文件

@ -52,7 +52,9 @@ def update_policy(
if grad_clip_norm > 0:
grad_norm = accelerator.clip_grad_norm_(policy.parameters(), grad_clip_norm)
else:
grad_norm = torch.nn.utils.clip_grad_norm_(policy.parameters(), float("inf"), error_if_nonfinite=False)
grad_norm = torch.nn.utils.clip_grad_norm_(
policy.parameters(), float("inf"), error_if_nonfinite=False
)
with lock if lock is not None else nullcontext():
optimizer.step()

查看文件

@ -14,7 +14,6 @@
"""Human-in-loop recording helpers used by `lerobot_record.py`."""
import logging
from collections import deque
from concurrent.futures import ThreadPoolExecutor
from copy import deepcopy
@ -53,7 +52,9 @@ def _get_torch_rng_state(device: torch.device) -> tuple[torch.Tensor, torch.Tens
return cpu_state, cuda_state
def _set_torch_rng_state(device: torch.device, cpu_state: torch.Tensor, cuda_state: torch.Tensor | None) -> None:
def _set_torch_rng_state(
device: torch.device, cpu_state: torch.Tensor, cuda_state: torch.Tensor | None
) -> None:
torch.set_rng_state(cpu_state)
if device.type == "cuda" and cuda_state is not None:
torch.cuda.set_rng_state(cuda_state, device)

查看文件

@ -35,10 +35,10 @@ from lerobot.processor import (
)
from lerobot.robots import Robot
from lerobot.scripts.recording_hil import (
ACPInferenceConfig,
INTERVENTION_STATE_ACTIVE,
INTERVENTION_STATE_POLICY,
INTERVENTION_STATE_RELEASE,
ACPInferenceConfig,
PolicySyncDualArmExecutor,
_capture_policy_runtime_state,
_predict_policy_action_with_acp_inference,
@ -155,8 +155,8 @@ def record_loop(
action_feature_names = list(robot.action_features.keys())
else:
action_feature_names = list(robot.action_features)
zero_policy_action = {name: 0.0 for name in action_feature_names}
has_teleop = isinstance(teleop, Teleoperator) or isinstance(teleop, list)
zero_policy_action = dict.fromkeys(action_feature_names, 0.0)
has_teleop = isinstance(teleop, (Teleoperator, list))
intervention_enabled = intervention_state_machine_enabled and policy is not None and has_teleop
intervention_state = INTERVENTION_STATE_POLICY
last_teleop_action: RobotAction | None = None
@ -282,22 +282,26 @@ def record_loop(
# Get action from policy and/or teleop
act_processed_policy: RobotAction | None = None
act_processed_teleop: RobotAction | None = None
if policy is not None and preprocessor is not None and postprocessor is not None:
if not (intervention_enabled and intervention_state == INTERVENTION_STATE_ACTIVE):
policy_action = _predict_policy_action_with_acp_inference(
observation_frame=observation_frame,
policy=policy,
device=get_safe_torch_device(policy.config.device),
preprocessor=preprocessor,
postprocessor=postprocessor,
use_amp=policy.config.use_amp,
task=single_task,
robot_type=robot.robot_type,
acp_inference=acp_inference,
cond_runtime_state=cond_policy_runtime_state,
uncond_runtime_state=uncond_policy_runtime_state,
)
act_processed_policy = make_robot_action(policy_action, dataset.features)
if (
policy is not None
and preprocessor is not None
and postprocessor is not None
and not (intervention_enabled and intervention_state == INTERVENTION_STATE_ACTIVE)
):
policy_action = _predict_policy_action_with_acp_inference(
observation_frame=observation_frame,
policy=policy,
device=get_safe_torch_device(policy.config.device),
preprocessor=preprocessor,
postprocessor=postprocessor,
use_amp=policy.config.use_amp,
task=single_task,
robot_type=robot.robot_type,
acp_inference=acp_inference,
cond_runtime_state=cond_policy_runtime_state,
uncond_runtime_state=uncond_policy_runtime_state,
)
act_processed_policy = make_robot_action(policy_action, dataset.features)
if isinstance(teleop, Teleoperator):
act = run_with_connection_retry("teleop.get_action", teleop.get_action)
@ -325,7 +329,9 @@ def record_loop(
last_teleop_action = act_processed_teleop
teleop_fallback_warned = False
policy_action_for_storage = act_processed_policy if act_processed_policy is not None else zero_policy_action
policy_action_for_storage = (
act_processed_policy if act_processed_policy is not None else zero_policy_action
)
is_intervention = 0.0
if intervention_enabled and intervention_state == INTERVENTION_STATE_ACTIVE:
@ -367,12 +373,14 @@ def record_loop(
if policy_sync_executor is not None and selected_from_policy:
_sent_action = run_with_connection_retry(
"policy_sync_executor.send_action",
lambda: policy_sync_executor.send_action(robot_action_to_send),
lambda robot_action_to_send=robot_action_to_send: policy_sync_executor.send_action(
robot_action_to_send
),
)
else:
_sent_action = run_with_connection_retry(
"robot.send_action",
lambda: robot.send_action(robot_action_to_send),
lambda robot_action_to_send=robot_action_to_send: robot.send_action(robot_action_to_send),
)
# Write to dataset

查看文件

@ -1,14 +1,14 @@
#!/usr/bin/env python
from pathlib import Path
import tempfile
from pathlib import Path
import numpy as np
from PIL import Image, ImageDraw, ImageFont
from tqdm.auto import tqdm
from lerobot.datasets.video_utils import decode_video_frames, encode_video_frames
from lerobot.datasets.lerobot_dataset import LeRobotDataset
from lerobot.datasets.video_utils import decode_video_frames, encode_video_frames
def _select_video_key(camera_keys: list[str], requested_video_key: str | None) -> str:
@ -16,7 +16,9 @@ def _select_video_key(camera_keys: list[str], requested_video_key: str | None) -
raise ValueError("No camera key found in dataset.")
if requested_video_key is not None:
if requested_video_key not in camera_keys:
raise ValueError(f"Unknown video_key '{requested_video_key}'. Available camera keys: {camera_keys}")
raise ValueError(
f"Unknown video_key '{requested_video_key}'. Available camera keys: {camera_keys}"
)
return requested_video_key
for key in camera_keys:
@ -277,7 +279,7 @@ def _encode_pil_to_video(
fps: int,
vcodec: str,
) -> None:
"""Encode PIL frames directly to video without writing intermediate PNGs."""
"""Encode PIL frames directly to video without writing intermediate PNG files."""
import av as _av
video_options, pix_fmt = _get_video_encode_options(vcodec)
@ -456,7 +458,9 @@ def _export_overlay_videos(
)
written_paths: list[Path] = []
for src, dst, vals, advs, inds, ts in tqdm(tasks, total=len(tasks), desc="Export value overlay videos", leave=False):
for src, dst, vals, advs, inds, ts in tqdm(
tasks, total=len(tasks), desc="Export value overlay videos", leave=False
):
written_paths.append(
_export_single_episode(
src,

查看文件

@ -116,15 +116,11 @@ class BiSOLeader(Teleoperator):
def send_feedback(self, feedback: dict[str, float]) -> None:
# Remove "left_" prefix
left_feedback = {
key.removeprefix("left_"): value
for key, value in feedback.items()
if key.startswith("left_")
key.removeprefix("left_"): value for key, value in feedback.items() if key.startswith("left_")
}
# Remove "right_" prefix
right_feedback = {
key.removeprefix("right_"): value
for key, value in feedback.items()
if key.startswith("right_")
key.removeprefix("right_"): value for key, value in feedback.items() if key.startswith("right_")
}
self.left_arm.send_feedback(left_feedback)

查看文件

@ -165,9 +165,7 @@ class SOLeader(Teleoperator):
# For phase-A dual-arm execution, reuse leader as a commanded arm by
# mapping `{joint}.pos` to motor goal positions.
self.set_manual_control(False)
goal_pos = {
key.removesuffix(".pos"): val for key, val in feedback.items() if key.endswith(".pos")
}
goal_pos = {key.removesuffix(".pos"): val for key, val in feedback.items() if key.endswith(".pos")}
if not goal_pos:
return
self.bus.sync_write("Goal_Position", goal_pos)

查看文件

@ -103,4 +103,3 @@ def resolve_collector_policy_id(
if intervention_enabled:
return human_id if is_intervention else policy_id
return policy_id if selected_from_policy else human_id

查看文件

@ -1 +1 @@
#!/usr/bin/env python
#!/usr/bin/env python

查看文件

@ -12,7 +12,7 @@ from typing import TYPE_CHECKING, Any
import numpy as np
import torch
import torch.nn.functional as F
import torch.nn.functional as functional
from huggingface_hub import hf_hub_download
from huggingface_hub.constants import SAFETENSORS_SINGLE_FILE
from huggingface_hub.errors import HfHubHTTPError
@ -78,7 +78,7 @@ def project_values_to_bins(values: torch.Tensor, bin_centers: torch.Tensor) -> t
def expected_value_from_logits(logits: torch.Tensor, bin_centers: torch.Tensor) -> torch.Tensor:
probs = F.softmax(logits, dim=-1)
probs = functional.softmax(logits, dim=-1)
return (probs * bin_centers).sum(dim=-1)
@ -198,13 +198,13 @@ def _resolve_image_size(image_processor: Any) -> tuple[int, int]:
return 384, 384
def _resolve_norm_stats(image_processor: Any) -> tuple[tuple[float, float, float], tuple[float, float, float]]:
def _resolve_norm_stats(
image_processor: Any,
) -> tuple[tuple[float, float, float], tuple[float, float, float]]:
mean_raw = getattr(image_processor, "image_mean", [0.5, 0.5, 0.5])
std_raw = getattr(image_processor, "image_std", [0.5, 0.5, 0.5])
if len(mean_raw) != 3 or len(std_raw) != 3:
raise ValueError(
f"Expected RGB normalization stats of len=3, got mean={mean_raw} std={std_raw}."
)
raise ValueError(f"Expected RGB normalization stats of len=3, got mean={mean_raw} std={std_raw}.")
mean = (float(mean_raw[0]), float(mean_raw[1]), float(mean_raw[2]))
std = (float(std_raw[0]), float(std_raw[1]), float(std_raw[2]))
if any(v <= 0 for v in std):
@ -367,7 +367,7 @@ class Pistar06Model(nn.Module):
flat_images = images.view(bsize * num_cameras, *images.shape[2:])
if flat_images.shape[-2:] != self.image_resolution:
flat_images = F.interpolate(
flat_images = functional.interpolate(
flat_images,
size=self.image_resolution,
mode="bilinear",
@ -435,7 +435,9 @@ class Pistar06Model(nn.Module):
language_context = torch.no_grad() if self.cfg.freeze_language_model else nullcontext()
with language_context:
language_features = self._encode_language(input_ids=input_ids, attention_mask=language_mask.long())
language_features = self._encode_language(
input_ids=input_ids, attention_mask=language_mask.long()
)
feature_dtype = torch.float32
image_features = image_features.to(dtype=feature_dtype)
@ -445,7 +447,9 @@ class Pistar06Model(nn.Module):
camera_token_mask = image_attention_mask.unsqueeze(-1).to(dtype=image_tokens.dtype)
image_tokens = image_tokens * camera_token_mask
camera_denominator = image_attention_mask.sum(dim=1, keepdim=True).to(dtype=image_tokens.dtype).clamp_min(1.0)
camera_denominator = (
image_attention_mask.sum(dim=1, keepdim=True).to(dtype=image_tokens.dtype).clamp_min(1.0)
)
image_pooled = image_tokens.sum(dim=1) / camera_denominator
language_token = self.language_projector(language_features)
@ -743,7 +747,7 @@ class Pistar06Policy(PreTrainedPolicy):
bin_centers = self.bin_centers.to(device=device)
soft_target = project_values_to_bins(value_target, bin_centers)
log_probs = F.log_softmax(logits, dim=-1)
log_probs = functional.log_softmax(logits, dim=-1)
per_sample_loss = -(soft_target * log_probs).sum(dim=-1)
sample_weight = None
@ -769,13 +773,12 @@ class Pistar06Policy(PreTrainedPolicy):
pred_value = expected_value_from_logits(logits, bin_centers)
value_mae = (pred_value - value_target).abs().mean()
if reduction == "none":
loss = per_sample_loss
else:
loss = per_sample_loss.mean()
loss = per_sample_loss if reduction == "none" else per_sample_loss.mean()
loss_dict = {
"loss": float(loss.mean().detach().item()) if reduction == "none" else float(loss.detach().item()),
"loss": float(loss.mean().detach().item())
if reduction == "none"
else float(loss.detach().item()),
"value_mae": float(value_mae.detach().item()),
}
if sample_weight is not None:

查看文件

@ -8,7 +8,7 @@ from typing import Any
import numpy as np
import torch
import torch.nn.functional as F
import torch.nn.functional as functional
from torch import Tensor
from lerobot.configs.types import PipelineFeatureType, PolicyFeature
@ -39,7 +39,7 @@ PISTAR06_IMAGE_MASK_KEY = "observation.pistar06.image_attention_mask"
def _pad_last_dim(vector: Tensor, new_dim: int) -> Tensor:
if vector.shape[-1] >= new_dim:
return vector
return F.pad(vector, (0, new_dim - vector.shape[-1]))
return functional.pad(vector, (0, new_dim - vector.shape[-1]))
@ProcessorStepRegistry.register(name="pistar06_prepare_task_prompt")
@ -216,7 +216,7 @@ def make_pistar06_pre_post_processors(
]:
camera_features = list(config.camera_features)
if not camera_features:
camera_features = [k for k in (config.input_features or {}).keys() if k.startswith(OBS_IMAGES)]
camera_features = [k for k in (config.input_features or {}) if k.startswith(OBS_IMAGES)]
input_steps: list[ProcessorStep] = [
RenameObservationsProcessorStep(rename_map={}),

查看文件

@ -729,7 +729,9 @@ def test_delete_episodes_preserves_extra_episode_metadata(tmp_path, empty_lerobo
)
assert "episode_success" in new_dataset.meta.episodes.column_names
assert [new_dataset.meta.episodes[i]["episode_success"] for i in range(new_dataset.meta.total_episodes)] == [
assert [
new_dataset.meta.episodes[i]["episode_success"] for i in range(new_dataset.meta.total_episodes)
] == [
"success",
"success",
]

查看文件

@ -15,9 +15,8 @@
# limitations under the License.
import json
from unittest.mock import MagicMock, patch
from collections import deque
from unittest.mock import MagicMock, patch
import numpy as np
import pytest
@ -25,12 +24,12 @@ import torch
from lerobot.datasets.lerobot_dataset import LeRobotDataset
from lerobot.scripts.lerobot_calibrate import CalibrateConfig, calibrate
from lerobot.scripts.lerobot_human_inloop_record import human_inloop_record
from lerobot.scripts.lerobot_human_inloop_record import (
_HumanInloopFailureResetController,
_load_failure_reset_pose,
_save_failure_reset_pose,
_slow_reset_all_arms_to_pose,
human_inloop_record,
)
from lerobot.scripts.lerobot_record import (
ACPInferenceConfig,
@ -39,8 +38,8 @@ from lerobot.scripts.lerobot_record import (
RecordConfig,
_capture_policy_runtime_state,
_predict_policy_action_with_acp_inference,
record_loop,
record,
record_loop,
)
from lerobot.scripts.lerobot_replay import DatasetReplayConfig, ReplayConfig, replay
from lerobot.scripts.lerobot_teleoperate import TeleoperateConfig, teleoperate
@ -211,9 +210,7 @@ def test_record_loop_sets_leader_manual_control_during_reset():
def test_save_and_load_failure_reset_pose(tmp_path):
robot = MockRobot(
MockRobotConfig(n_motors=2, random_values=False, static_values=[12.5, -3.0])
)
robot = MockRobot(MockRobotConfig(n_motors=2, random_values=False, static_values=[12.5, -3.0]))
robot.connect()
pose_path = tmp_path / "failure_reset_pose.json"

查看文件

@ -9,7 +9,9 @@ def test_contiguous_distributed_eval_sampler_has_contiguous_shards_with_tail_pad
shards: list[list[int]] = []
for rank in range(world_size):
sampler = ContiguousDistributedEvalSampler(dataset_size=dataset_size, num_replicas=world_size, rank=rank)
sampler = ContiguousDistributedEvalSampler(
dataset_size=dataset_size, num_replicas=world_size, rank=rank
)
shard = list(iter(sampler))
shards.append(shard)
assert len(shard) == len(sampler) == 4
@ -30,7 +32,9 @@ def test_contiguous_distributed_eval_sampler_handles_more_ranks_than_samples():
shards: list[list[int]] = []
for rank in range(world_size):
sampler = ContiguousDistributedEvalSampler(dataset_size=dataset_size, num_replicas=world_size, rank=rank)
sampler = ContiguousDistributedEvalSampler(
dataset_size=dataset_size, num_replicas=world_size, rank=rank
)
shard = list(iter(sampler))
shards.append(shard)
assert len(shard) == len(sampler) == 1

查看文件

@ -13,8 +13,7 @@ from lerobot.configs.value import (
ValueInferencePipelineConfig,
ValueInferenceVizConfig,
)
from lerobot.scripts import lerobot_value_infer as value_infer
from lerobot.scripts import value_infer_viz
from lerobot.scripts import lerobot_value_infer as value_infer, value_infer_viz
class _FakeHFDataset:

查看文件

@ -10,8 +10,6 @@ from torch import nn
import lerobot.processor.tokenizer_processor as tokenizer_processor
import lerobot.values.pistar06.modeling_pistar06 as pistar06_modeling
import lerobot.values.pistar06.processor_pistar06 as pistar06_processor
from lerobot.configs.value_train import ValueTargetsConfig
from lerobot.utils.constants import OBS_LANGUAGE_ATTENTION_MASK, OBS_LANGUAGE_TOKENS, OBS_STATE
from lerobot.values.pistar06.configuration_pistar06 import Pistar06Config
from lerobot.values.pistar06.modeling_pistar06 import (
@ -95,10 +93,7 @@ class _DummyAutoModel:
@classmethod
def from_pretrained(cls, repo_id, **kwargs):
output_loading_info = bool(kwargs.pop("output_loading_info", False))
if "siglip" in repo_id.lower():
model = _DummyVisionModel()
else:
model = _DummyLanguageModel()
model = _DummyVisionModel() if "siglip" in repo_id.lower() else _DummyLanguageModel()
if output_loading_info:
return model, {"missing_keys": [], "unexpected_keys": [], "mismatched_keys": []}
return model