Hardware API redesign (#777)

Co-authored-by: Pepijn <138571049+pkooij@users.noreply.github.com>
Co-authored-by: Steven Palma <imstevenpmwork@ieee.org>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Steven Palma <steven.palma@huggingface.co>
Co-authored-by: Adil Zouitine <adilzouitinegm@gmail.com>
Co-authored-by: Pepijn <pepijn@huggingface.co>
This commit is contained in:
Simon Alibert
2025-06-05 17:48:43 +02:00
committed by GitHub
parent b536f47e3f
commit e23b41e79a
247 changed files with 14442 additions and 13550 deletions
+17
View File
@@ -0,0 +1,17 @@
# Copyright 2024 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.
from .camera import Camera
from .configs import CameraConfig, ColorMode, Cv2Rotation
from .utils import make_cameras_from_configs
+120
View File
@@ -0,0 +1,120 @@
#!/usr/bin/env python
# Copyright 2024 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.
import abc
from typing import Any, Dict, List
import numpy as np
from .configs import CameraConfig, ColorMode
class Camera(abc.ABC):
"""Base class for camera implementations.
Defines a standard interface for camera operations across different backends.
Subclasses must implement all abstract methods.
Manages basic camera properties (FPS, resolution) and core operations:
- Connection/disconnection
- Frame capture (sync/async)
Attributes:
fps (int | None): Configured frames per second
width (int | None): Frame width in pixels
height (int | None): Frame height in pixels
Example:
class MyCamera(Camera):
def __init__(self, config): ...
@property
def is_connected(self) -> bool: ...
def connect(self, warmup=True): ...
# Plus other required methods
"""
def __init__(self, config: CameraConfig):
"""Initialize the camera with the given configuration.
Args:
config: Camera configuration containing FPS and resolution.
"""
self.fps: int | None = config.fps
self.width: int | None = config.width
self.height: int | None = config.height
@property
@abc.abstractmethod
def is_connected(self) -> bool:
"""Check if the camera is currently connected.
Returns:
bool: True if the camera is connected and ready to capture frames,
False otherwise.
"""
pass
@staticmethod
@abc.abstractmethod
def find_cameras() -> List[Dict[str, Any]]:
"""Detects available cameras connected to the system.
Returns:
List[Dict[str, Any]]: A list of dictionaries,
where each dictionary contains information about a detected camera.
"""
pass
@abc.abstractmethod
def connect(self, warmup: bool = True) -> None:
"""Establish connection to the camera.
Args:
warmup: If True (default), captures a warmup frame before returning. Useful
for cameras that require time to adjust capture settings.
If False, skips the warmup frame.
"""
pass
@abc.abstractmethod
def read(self, color_mode: ColorMode | None = None) -> np.ndarray:
"""Capture and return a single frame from the camera.
Args:
color_mode: Desired color mode for the output frame. If None,
uses the camera's default color mode.
Returns:
np.ndarray: Captured frame as a numpy array.
"""
pass
@abc.abstractmethod
def async_read(self, timeout_ms: float = ...) -> np.ndarray:
"""Asynchronously capture and return a single frame from the camera.
Args:
timeout_ms: Maximum time to wait for a frame in milliseconds.
Defaults to implementation-specific timeout.
Returns:
np.ndarray: Captured frame as a numpy array.
"""
pass
@abc.abstractmethod
def disconnect(self) -> None:
"""Disconnect from the camera and release resources."""
pass
+44
View File
@@ -0,0 +1,44 @@
#!/usr/bin/env python
# Copyright 2024 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.
import abc
from dataclasses import dataclass
from enum import Enum
import draccus
class ColorMode(str, Enum):
RGB = "rgb"
BGR = "bgr"
class Cv2Rotation(int, Enum):
NO_ROTATION = 0
ROTATE_90 = 90
ROTATE_180 = 180
ROTATE_270 = -90
@dataclass(kw_only=True)
class CameraConfig(draccus.ChoiceRegistry, abc.ABC):
fps: int | None = None
width: int | None = None
height: int | None = None
@property
def type(self) -> str:
return self.get_choice_name(self.__class__)
+16
View File
@@ -0,0 +1,16 @@
# Copyright 2024 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.
from .camera_opencv import OpenCVCamera
from .configuration_opencv import OpenCVCameraConfig
@@ -0,0 +1,479 @@
# Copyright 2024 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.
"""
Provides the OpenCVCamera class for capturing frames from cameras using OpenCV.
"""
import logging
import math
import platform
import time
from pathlib import Path
from threading import Event, Lock, Thread
from typing import Any, Dict, List
import cv2
import numpy as np
from lerobot.common.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
from ..camera import Camera
from ..utils import get_cv2_backend, get_cv2_rotation
from .configuration_opencv import ColorMode, OpenCVCameraConfig
# NOTE(Steven): The maximum opencv device index depends on your operating system. For instance,
# if you have 3 cameras, they should be associated to index 0, 1, and 2. This is the case
# on MacOS. However, on Ubuntu, the indices are different like 6, 16, 23.
# When you change the USB port or reboot the computer, the operating system might
# treat the same cameras as new devices. Thus we select a higher bound to search indices.
MAX_OPENCV_INDEX = 60
logger = logging.getLogger(__name__)
class OpenCVCamera(Camera):
"""
Manages camera interactions using OpenCV for efficient frame recording.
This class provides a high-level interface to connect to, configure, and read
frames from cameras compatible with OpenCV's VideoCapture. It supports both
synchronous and asynchronous frame reading.
An OpenCVCamera instance requires a camera index (e.g., 0) or a device path
(e.g., '/dev/video0' on Linux). Camera indices can be unstable across reboots
or port changes, especially on Linux. Use the provided utility script to find
available camera indices or paths:
```bash
python -m lerobot.find_cameras opencv
```
The camera's default settings (FPS, resolution, color mode) are used unless
overridden in the configuration.
Example:
```python
from lerobot.common.cameras.opencv import OpenCVCamera
from lerobot.common.cameras.configuration_opencv import OpenCVCameraConfig, ColorMode, Cv2Rotation
# Basic usage with camera index 0
config = OpenCVCameraConfig(index_or_path=0)
camera = OpenCVCamera(config)
camera.connect()
# Read 1 frame synchronously
color_image = camera.read()
print(color_image.shape)
# Read 1 frame asynchronously
async_image = camera.async_read()
# When done, properly disconnect the camera using
camera.disconnect()
# Example with custom settings
custom_config = OpenCVCameraConfig(
index_or_path='/dev/video0', # Or use an index
fps=30,
width=1280,
height=720,
color_mode=ColorMode.RGB,
rotation=Cv2Rotation.ROTATE_90
)
custom_camera = OpenCVCamera(custom_config)
# ... connect, read, disconnect ...
```
"""
def __init__(self, config: OpenCVCameraConfig):
"""
Initializes the OpenCVCamera instance.
Args:
config: The configuration settings for the camera.
"""
super().__init__(config)
self.config = config
self.index_or_path = config.index_or_path
self.fps = config.fps
self.color_mode = config.color_mode
self.warmup_s = config.warmup_s
self.videocapture: cv2.VideoCapture | None = None
self.thread: Thread | None = None
self.stop_event: Event | None = None
self.frame_lock: Lock = Lock()
self.latest_frame: np.ndarray | None = None
self.new_frame_event: Event = Event()
self.rotation: int | None = get_cv2_rotation(config.rotation)
self.backend: int = get_cv2_backend()
if self.height and self.width:
self.capture_width, self.capture_height = self.width, self.height
if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
self.capture_width, self.capture_height = self.height, self.width
def __str__(self) -> str:
return f"{self.__class__.__name__}({self.index_or_path})"
@property
def is_connected(self) -> bool:
"""Checks if the camera is currently connected and opened."""
return isinstance(self.videocapture, cv2.VideoCapture) and self.videocapture.isOpened()
def connect(self, warmup: bool = True):
"""
Connects to the OpenCV camera specified in the configuration.
Initializes the OpenCV VideoCapture object, sets desired camera properties
(FPS, width, height), and performs initial checks.
Raises:
DeviceAlreadyConnectedError: If the camera is already connected.
ConnectionError: If the specified camera index/path is not found or the camera is found but fails to open.
RuntimeError: If the camera opens but fails to apply requested FPS/resolution settings.
"""
if self.is_connected:
raise DeviceAlreadyConnectedError(f"{self} is already connected.")
# Use 1 thread for OpenCV operations to avoid potential conflicts or
# blocking in multi-threaded applications, especially during data collection.
cv2.setNumThreads(1)
self.videocapture = cv2.VideoCapture(self.index_or_path, self.backend)
if not self.videocapture.isOpened():
self.videocapture.release()
self.videocapture = None
raise ConnectionError(
f"Failed to open {self}."
f"Run `python -m lerobot.find_cameras opencv` to find available cameras."
)
self._configure_capture_settings()
if warmup:
start_time = time.time()
while time.time() - start_time < self.warmup_s:
self.read()
time.sleep(0.1)
logger.info(f"{self} connected.")
def _configure_capture_settings(self) -> None:
"""
Applies the specified FPS, width, and height settings to the connected camera.
This method attempts to set the camera properties via OpenCV. It checks if
the camera successfully applied the settings and raises an error if not.
Args:
fps: The desired frames per second. If None, the setting is skipped.
width: The desired capture width. If None, the setting is skipped.
height: The desired capture height. If None, the setting is skipped.
Raises:
RuntimeError: If the camera fails to set any of the specified properties
to the requested value.
DeviceNotConnectedError: If the camera is not connected when attempting
to configure settings.
"""
if not self.is_connected:
raise DeviceNotConnectedError(f"Cannot configure settings for {self} as it is not connected.")
if self.fps is None:
self.fps = self.videocapture.get(cv2.CAP_PROP_FPS)
else:
self._validate_fps()
default_width = int(round(self.videocapture.get(cv2.CAP_PROP_FRAME_WIDTH)))
default_height = int(round(self.videocapture.get(cv2.CAP_PROP_FRAME_HEIGHT)))
if self.width is None or self.height is None:
self.width, self.height = default_width, default_height
self.capture_width, self.capture_height = default_width, default_height
if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
self.width, self.height = default_height, default_width
self.capture_width, self.capture_height = default_width, default_height
else:
self._validate_width_and_height()
def _validate_fps(self) -> None:
"""Validates and sets the camera's frames per second (FPS)."""
success = self.videocapture.set(cv2.CAP_PROP_FPS, float(self.fps))
actual_fps = self.videocapture.get(cv2.CAP_PROP_FPS)
# Use math.isclose for robust float comparison
if not success or not math.isclose(self.fps, actual_fps, rel_tol=1e-3):
raise RuntimeError(f"{self} failed to set fps={self.fps} ({actual_fps=}).")
def _validate_width_and_height(self) -> None:
"""Validates and sets the camera's frame capture width and height."""
success = self.videocapture.set(cv2.CAP_PROP_FRAME_WIDTH, float(self.capture_width))
actual_width = int(round(self.videocapture.get(cv2.CAP_PROP_FRAME_WIDTH)))
if not success or self.capture_width != actual_width:
raise RuntimeError(f"{self} failed to set capture_width={self.capture_width} ({actual_width=}).")
success = self.videocapture.set(cv2.CAP_PROP_FRAME_HEIGHT, float(self.capture_height))
actual_height = int(round(self.videocapture.get(cv2.CAP_PROP_FRAME_HEIGHT)))
if not success or self.capture_height != actual_height:
raise RuntimeError(
f"{self} failed to set capture_height={self.capture_height} ({actual_height})."
)
@staticmethod
def find_cameras() -> List[Dict[str, Any]]:
"""
Detects available OpenCV cameras connected to the system.
On Linux, it scans '/dev/video*' paths. On other systems (like macOS, Windows),
it checks indices from 0 up to `MAX_OPENCV_INDEX`.
Returns:
List[Dict[str, Any]]: A list of dictionaries,
where each dictionary contains 'type', 'id' (port index or path),
and the default profile properties (width, height, fps, format).
"""
found_cameras_info = []
if platform.system() == "Linux":
possible_paths = sorted(Path("/dev").glob("video*"), key=lambda p: p.name)
targets_to_scan = [str(p) for p in possible_paths]
else:
targets_to_scan = list(range(MAX_OPENCV_INDEX))
for target in targets_to_scan:
camera = cv2.VideoCapture(target)
if camera.isOpened():
default_width = int(camera.get(cv2.CAP_PROP_FRAME_WIDTH))
default_height = int(camera.get(cv2.CAP_PROP_FRAME_HEIGHT))
default_fps = camera.get(cv2.CAP_PROP_FPS)
default_format = camera.get(cv2.CAP_PROP_FORMAT)
camera_info = {
"name": f"OpenCV Camera @ {target}",
"type": "OpenCV",
"id": target,
"backend_api": camera.getBackendName(),
"default_stream_profile": {
"format": default_format,
"width": default_width,
"height": default_height,
"fps": default_fps,
},
}
found_cameras_info.append(camera_info)
camera.release()
return found_cameras_info
def read(self, color_mode: ColorMode | None = None) -> np.ndarray:
"""
Reads a single frame synchronously from the camera.
This is a blocking call. It waits for the next available frame from the
camera hardware via OpenCV.
Args:
color_mode (Optional[ColorMode]): If specified, overrides the default
color mode (`self.color_mode`) for this read operation (e.g.,
request RGB even if default is BGR).
Returns:
np.ndarray: The captured frame as a NumPy array in the format
(height, width, channels), using the specified or default
color mode and applying any configured rotation.
Raises:
DeviceNotConnectedError: If the camera is not connected.
RuntimeError: If reading the frame from the camera fails or if the
received frame dimensions don't match expectations before rotation.
ValueError: If an invalid `color_mode` is requested.
"""
if not self.is_connected:
raise DeviceNotConnectedError(f"{self} is not connected.")
start_time = time.perf_counter()
ret, frame = self.videocapture.read()
if not ret or frame is None:
raise RuntimeError(f"{self} read failed (status={ret}).")
processed_frame = self._postprocess_image(frame, color_mode)
read_duration_ms = (time.perf_counter() - start_time) * 1e3
logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
return processed_frame
def _postprocess_image(self, image: np.ndarray, color_mode: ColorMode | None = None) -> np.ndarray:
"""
Applies color conversion, dimension validation, and rotation to a raw frame.
Args:
image (np.ndarray): The raw image frame (expected BGR format from OpenCV).
color_mode (Optional[ColorMode]): The target color mode (RGB or BGR). If None,
uses the instance's default `self.color_mode`.
Returns:
np.ndarray: The processed image frame.
Raises:
ValueError: If the requested `color_mode` is invalid.
RuntimeError: If the raw frame dimensions do not match the configured
`width` and `height`.
"""
requested_color_mode = self.color_mode if color_mode is None else color_mode
if requested_color_mode not in (ColorMode.RGB, ColorMode.BGR):
raise ValueError(
f"Invalid color mode '{requested_color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
)
h, w, c = image.shape
if h != self.capture_height or w != self.capture_width:
raise RuntimeError(
f"{self} frame width={w} or height={h} do not match configured width={self.capture_width} or height={self.capture_height}."
)
if c != 3:
raise RuntimeError(f"{self} frame channels={c} do not match expected 3 channels (RGB/BGR).")
processed_image = image
if requested_color_mode == ColorMode.RGB:
processed_image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
processed_image = cv2.rotate(processed_image, self.rotation)
return processed_image
def _read_loop(self):
"""
Internal loop run by the background thread for asynchronous reading.
On each iteration:
1. Reads a color frame
2. Stores result in latest_frame (thread-safe)
3. Sets new_frame_event to notify listeners
Stops on DeviceNotConnectedError, logs other errors and continues.
"""
while not self.stop_event.is_set():
try:
color_image = self.read()
with self.frame_lock:
self.latest_frame = color_image
self.new_frame_event.set()
except DeviceNotConnectedError:
break
except Exception as e:
logger.warning(f"Error reading frame in background thread for {self}: {e}")
def _start_read_thread(self) -> None:
"""Starts or restarts the background read thread if it's not running."""
if self.thread is not None and self.thread.is_alive():
self.thread.join(timeout=0.1)
if self.stop_event is not None:
self.stop_event.set()
self.stop_event = Event()
self.thread = Thread(target=self._read_loop, args=(), name=f"{self}_read_loop")
self.thread.daemon = True
self.thread.start()
def _stop_read_thread(self) -> None:
"""Signals the background read thread to stop and waits for it to join."""
if self.stop_event is not None:
self.stop_event.set()
if self.thread is not None and self.thread.is_alive():
self.thread.join(timeout=2.0)
self.thread = None
self.stop_event = None
def async_read(self, timeout_ms: float = 200) -> np.ndarray:
"""
Reads the latest available frame asynchronously.
This method retrieves the most recent frame captured by the background
read thread. It does not block waiting for the camera hardware directly,
but may wait up to timeout_ms for the background thread to provide a frame.
Args:
timeout_ms (float): Maximum time in milliseconds to wait for a frame
to become available. Defaults to 200ms (0.2 seconds).
Returns:
np.ndarray: The latest captured frame as a NumPy array in the format
(height, width, channels), processed according to configuration.
Raises:
DeviceNotConnectedError: If the camera is not connected.
TimeoutError: If no frame becomes available within the specified timeout.
RuntimeError: If an unexpected error occurs.
"""
if not self.is_connected:
raise DeviceNotConnectedError(f"{self} is not connected.")
if self.thread is None or not self.thread.is_alive():
self._start_read_thread()
if not self.new_frame_event.wait(timeout=timeout_ms / 1000.0):
thread_alive = self.thread is not None and self.thread.is_alive()
raise TimeoutError(
f"Timed out waiting for frame from camera {self} after {timeout_ms} ms. "
f"Read thread alive: {thread_alive}."
)
with self.frame_lock:
frame = self.latest_frame
self.new_frame_event.clear()
if frame is None:
raise RuntimeError(f"Internal error: Event set but no frame available for {self}.")
return frame
def disconnect(self):
"""
Disconnects from the camera and cleans up resources.
Stops the background read thread (if running) and releases the OpenCV
VideoCapture object.
Raises:
DeviceNotConnectedError: If the camera is already disconnected.
"""
if not self.is_connected and self.thread is None:
raise DeviceNotConnectedError(f"{self} not connected.")
if self.thread is not None:
self._stop_read_thread()
if self.videocapture is not None:
self.videocapture.release()
self.videocapture = None
logger.info(f"{self} disconnected.")
@@ -0,0 +1,73 @@
# Copyright 2024 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.
from dataclasses import dataclass
from pathlib import Path
from ..configs import CameraConfig, ColorMode, Cv2Rotation
@CameraConfig.register_subclass("opencv")
@dataclass
class OpenCVCameraConfig(CameraConfig):
"""Configuration class for OpenCV-based camera devices or video files.
This class provides configuration options for cameras accessed through OpenCV,
supporting both physical camera devices and video files. It includes settings
for resolution, frame rate, color mode, and image rotation.
Example configurations:
```python
# Basic configurations
OpenCVCameraConfig(0, 30, 1280, 720) # 1280x720 @ 30FPS
OpenCVCameraConfig(/dev/video4, 60, 640, 480) # 640x480 @ 60FPS
# Advanced configurations
OpenCVCameraConfig(128422271347, 30, 640, 480, rotation=Cv2Rotation.ROTATE_90) # With 90° rotation
```
Attributes:
index_or_path: Either an integer representing the camera device index,
or a Path object pointing to a video file.
fps: Requested frames per second for the color stream.
width: Requested frame width in pixels for the color stream.
height: Requested frame height in pixels for the color stream.
color_mode: Color mode for image output (RGB or BGR). Defaults to RGB.
rotation: Image rotation setting (0°, 90°, 180°, or 270°). Defaults to no rotation.
warmup_s: Time reading frames before returning from connect (in seconds)
Note:
- Only 3-channel color output (RGB/BGR) is currently supported.
"""
index_or_path: int | Path
color_mode: ColorMode = ColorMode.RGB
rotation: Cv2Rotation = Cv2Rotation.NO_ROTATION
warmup_s: int = 1
def __post_init__(self):
if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
raise ValueError(
f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."
)
if self.rotation not in (
Cv2Rotation.NO_ROTATION,
Cv2Rotation.ROTATE_90,
Cv2Rotation.ROTATE_180,
Cv2Rotation.ROTATE_270,
):
raise ValueError(
f"`rotation` is expected to be in {(Cv2Rotation.NO_ROTATION, Cv2Rotation.ROTATE_90, Cv2Rotation.ROTATE_180, Cv2Rotation.ROTATE_270)}, but {self.rotation} is provided."
)
@@ -0,0 +1,16 @@
# Copyright 2024 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.
from .camera_realsense import RealSenseCamera
from .configuration_realsense import RealSenseCameraConfig
@@ -0,0 +1,556 @@
# Copyright 2024 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.
"""
Provides the RealSenseCamera class for capturing frames from Intel RealSense cameras.
"""
import logging
import time
from threading import Event, Lock, Thread
from typing import Any, Dict, List
import cv2
import numpy as np
try:
import pyrealsense2 as rs
except Exception as e:
logging.info(f"Could not import realsense: {e}")
from lerobot.common.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
from ..camera import Camera
from ..configs import ColorMode
from ..utils import get_cv2_rotation
from .configuration_realsense import RealSenseCameraConfig
logger = logging.getLogger(__name__)
class RealSenseCamera(Camera):
"""
Manages interactions with Intel RealSense cameras for frame and depth recording.
This class provides an interface similar to `OpenCVCamera` but tailored for
RealSense devices, leveraging the `pyrealsense2` library. It uses the camera's
unique serial number for identification, offering more stability than device
indices, especially on Linux. It also supports capturing depth maps alongside
color frames.
Use the provided utility script to find available camera indices and default profiles:
```bash
python -m lerobot.find_cameras realsense
```
A `RealSenseCamera` instance requires a configuration object specifying the
camera's serial number or a unique device name. If using the name, ensure only
one camera with that name is connected.
The camera's default settings (FPS, resolution, color mode) from the stream
profile are used unless overridden in the configuration.
Example:
```python
from lerobot.common.cameras.realsense import RealSenseCamera, RealSenseCameraConfig
from lerobot.common.cameras import ColorMode, Cv2Rotation
# Basic usage with serial number
config = RealSenseCameraConfig(serial_number_or_name="0123456789") # Replace with actual SN
camera = RealSenseCamera(config)
camera.connect()
# Read 1 frame synchronously
color_image = camera.read()
print(color_image.shape)
# Read 1 frame asynchronously
async_image = camera.async_read()
# When done, properly disconnect the camera using
camera.disconnect()
# Example with depth capture and custom settings
custom_config = RealSenseCameraConfig(
serial_number_or_name="0123456789", # Replace with actual SN
fps=30,
width=1280,
height=720,
color_mode=ColorMode.BGR, # Request BGR output
rotation=Cv2Rotation.NO_ROTATION,
use_depth=True
)
depth_camera = RealSenseCamera(custom_config)
depth_camera.connect()
# Read 1 depth frame
depth_map = depth_camera.read_depth()
# Example using a unique camera name
name_config = RealSenseCameraConfig(serial_number_or_name="Intel RealSense D435") # If unique
name_camera = RealSenseCamera(name_config)
# ... connect, read, disconnect ...
```
"""
def __init__(self, config: RealSenseCameraConfig):
"""
Initializes the RealSenseCamera instance.
Args:
config: The configuration settings for the camera.
"""
super().__init__(config)
self.config = config
if config.serial_number_or_name.isdigit():
self.serial_number = config.serial_number_or_name
else:
self.serial_number = self._find_serial_number_from_name(config.serial_number_or_name)
self.fps = config.fps
self.color_mode = config.color_mode
self.use_depth = config.use_depth
self.warmup_s = config.warmup_s
self.rs_pipeline: rs.pipeline | None = None
self.rs_profile: rs.pipeline_profile | None = None
self.thread: Thread | None = None
self.stop_event: Event | None = None
self.frame_lock: Lock = Lock()
self.latest_frame: np.ndarray | None = None
self.new_frame_event: Event = Event()
self.rotation: int | None = get_cv2_rotation(config.rotation)
if self.height and self.width:
self.capture_width, self.capture_height = self.width, self.height
if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
self.capture_width, self.capture_height = self.height, self.width
def __str__(self) -> str:
return f"{self.__class__.__name__}({self.serial_number})"
@property
def is_connected(self) -> bool:
"""Checks if the camera pipeline is started and streams are active."""
return self.rs_pipeline is not None and self.rs_profile is not None
def connect(self, warmup: bool = True):
"""
Connects to the RealSense camera specified in the configuration.
Initializes the RealSense pipeline, configures the required streams (color
and optionally depth), starts the pipeline, and validates the actual stream settings.
Raises:
DeviceAlreadyConnectedError: If the camera is already connected.
ValueError: If the configuration is invalid (e.g., missing serial/name, name not unique).
ConnectionError: If the camera is found but fails to start the pipeline or no RealSense devices are detected at all.
RuntimeError: If the pipeline starts but fails to apply requested settings.
"""
if self.is_connected:
raise DeviceAlreadyConnectedError(f"{self} is already connected.")
self.rs_pipeline = rs.pipeline()
rs_config = rs.config()
self._configure_rs_pipeline_config(rs_config)
try:
self.rs_profile = self.rs_pipeline.start(rs_config)
except RuntimeError as e:
self.rs_profile = None
self.rs_pipeline = None
raise ConnectionError(
f"Failed to open {self}."
"Run `python -m lerobot.find_cameras realsense` to find available cameras."
) from e
self._configure_capture_settings()
if warmup:
time.sleep(
1
) # NOTE(Steven): RS cameras need a bit of time to warm up before the first read. If we don't wait, the first read from the warmup will raise.
start_time = time.time()
while time.time() - start_time < self.warmup_s:
self.read()
time.sleep(0.1)
logger.info(f"{self} connected.")
@staticmethod
def find_cameras() -> List[Dict[str, Any]]:
"""
Detects available Intel RealSense cameras connected to the system.
Returns:
List[Dict[str, Any]]: A list of dictionaries,
where each dictionary contains 'type', 'id' (serial number), 'name',
firmware version, USB type, and other available specs, and the default profile properties (width, height, fps, format).
Raises:
OSError: If pyrealsense2 is not installed.
ImportError: If pyrealsense2 is not installed.
"""
found_cameras_info = []
context = rs.context()
devices = context.query_devices()
for device in devices:
camera_info = {
"name": device.get_info(rs.camera_info.name),
"type": "RealSense",
"id": device.get_info(rs.camera_info.serial_number),
"firmware_version": device.get_info(rs.camera_info.firmware_version),
"usb_type_descriptor": device.get_info(rs.camera_info.usb_type_descriptor),
"physical_port": device.get_info(rs.camera_info.physical_port),
"product_id": device.get_info(rs.camera_info.product_id),
"product_line": device.get_info(rs.camera_info.product_line),
}
# Get stream profiles for each sensor
sensors = device.query_sensors()
for sensor in sensors:
profiles = sensor.get_stream_profiles()
for profile in profiles:
if profile.is_video_stream_profile() and profile.is_default():
vprofile = profile.as_video_stream_profile()
stream_info = {
"stream_type": vprofile.stream_name(),
"format": vprofile.format().name,
"width": vprofile.width(),
"height": vprofile.height(),
"fps": vprofile.fps(),
}
camera_info["default_stream_profile"] = stream_info
found_cameras_info.append(camera_info)
return found_cameras_info
def _find_serial_number_from_name(self, name: str) -> str:
"""Finds the serial number for a given unique camera name."""
camera_infos = self.find_cameras()
found_devices = [cam for cam in camera_infos if str(cam["name"]) == name]
if not found_devices:
available_names = [cam["name"] for cam in camera_infos]
raise ValueError(
f"No RealSense camera found with name '{name}'. Available camera names: {available_names}"
)
if len(found_devices) > 1:
serial_numbers = [dev["serial_number"] for dev in found_devices]
raise ValueError(
f"Multiple RealSense cameras found with name '{name}'. "
f"Please use a unique serial number instead. Found SNs: {serial_numbers}"
)
serial_number = str(found_devices[0]["serial_number"])
return serial_number
def _configure_rs_pipeline_config(self, rs_config):
"""Creates and configures the RealSense pipeline configuration object."""
rs.config.enable_device(rs_config, self.serial_number)
if self.width and self.height and self.fps:
rs_config.enable_stream(
rs.stream.color, self.capture_width, self.capture_height, rs.format.rgb8, self.fps
)
if self.use_depth:
rs_config.enable_stream(
rs.stream.depth, self.capture_width, self.capture_height, rs.format.z16, self.fps
)
else:
rs_config.enable_stream(rs.stream.color)
if self.use_depth:
rs_config.enable_stream(rs.stream.depth)
def _configure_capture_settings(self) -> None:
"""Sets fps, width, and height from device stream if not already configured.
Uses the color stream profile to update unset attributes. Handles rotation by
swapping width/height when needed. Original capture dimensions are always stored.
Raises:
DeviceNotConnectedError: If device is not connected.
"""
if not self.is_connected:
raise DeviceNotConnectedError(f"Cannot validate settings for {self} as it is not connected.")
stream = self.rs_profile.get_stream(rs.stream.color).as_video_stream_profile()
if self.fps is None:
self.fps = stream.fps()
if self.width is None or self.height is None:
actual_width = int(round(stream.width()))
actual_height = int(round(stream.height()))
if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
self.width, self.height = actual_height, actual_width
self.capture_width, self.capture_height = actual_width, actual_height
else:
self.width, self.height = actual_width, actual_height
self.capture_width, self.capture_height = actual_width, actual_height
def read_depth(self, timeout_ms: int = 200) -> np.ndarray:
"""
Reads a single frame (depth) synchronously from the camera.
This is a blocking call. It waits for a coherent set of frames (depth)
from the camera hardware via the RealSense pipeline.
Args:
timeout_ms (int): Maximum time in milliseconds to wait for a frame. Defaults to 200ms.
Returns:
np.ndarray: The depth map as a NumPy array (height, width)
of type `np.uint16` (raw depth values in millimeters) and rotation.
Raises:
DeviceNotConnectedError: If the camera is not connected.
RuntimeError: If reading frames from the pipeline fails or frames are invalid.
"""
if not self.is_connected:
raise DeviceNotConnectedError(f"{self} is not connected.")
if not self.use_depth:
raise RuntimeError(
f"Failed to capture depth frame '.read_depth()'. Depth stream is not enabled for {self}."
)
start_time = time.perf_counter()
ret, frame = self.rs_pipeline.try_wait_for_frames(timeout_ms=timeout_ms)
if not ret or frame is None:
raise RuntimeError(f"{self} read_depth failed (status={ret}).")
depth_frame = frame.get_depth_frame()
depth_map = np.asanyarray(depth_frame.get_data())
depth_map_processed = self._postprocess_image(depth_map, depth_frame=True)
read_duration_ms = (time.perf_counter() - start_time) * 1e3
logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
return depth_map_processed
def read(self, color_mode: ColorMode | None = None, timeout_ms: int = 200) -> np.ndarray:
"""
Reads a single frame (color) synchronously from the camera.
This is a blocking call. It waits for a coherent set of frames (color)
from the camera hardware via the RealSense pipeline.
Args:
timeout_ms (int): Maximum time in milliseconds to wait for a frame. Defaults to 200ms.
Returns:
np.ndarray: The captured color frame as a NumPy array
(height, width, channels), processed according to `color_mode` and rotation.
Raises:
DeviceNotConnectedError: If the camera is not connected.
RuntimeError: If reading frames from the pipeline fails or frames are invalid.
ValueError: If an invalid `color_mode` is requested.
"""
if not self.is_connected:
raise DeviceNotConnectedError(f"{self} is not connected.")
start_time = time.perf_counter()
ret, frame = self.rs_pipeline.try_wait_for_frames(timeout_ms=timeout_ms)
if not ret or frame is None:
raise RuntimeError(f"{self} read failed (status={ret}).")
color_frame = frame.get_color_frame()
color_image_raw = np.asanyarray(color_frame.get_data())
color_image_processed = self._postprocess_image(color_image_raw, color_mode)
read_duration_ms = (time.perf_counter() - start_time) * 1e3
logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
return color_image_processed
def _postprocess_image(
self, image: np.ndarray, color_mode: ColorMode | None = None, depth_frame: bool = False
) -> np.ndarray:
"""
Applies color conversion, dimension validation, and rotation to a raw color frame.
Args:
image (np.ndarray): The raw image frame (expected RGB format from RealSense).
color_mode (Optional[ColorMode]): The target color mode (RGB or BGR). If None,
uses the instance's default `self.color_mode`.
Returns:
np.ndarray: The processed image frame according to `self.color_mode` and `self.rotation`.
Raises:
ValueError: If the requested `color_mode` is invalid.
RuntimeError: If the raw frame dimensions do not match the configured
`width` and `height`.
"""
if color_mode and color_mode not in (ColorMode.RGB, ColorMode.BGR):
raise ValueError(
f"Invalid requested color mode '{color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
)
if depth_frame:
h, w = image.shape
else:
h, w, c = image.shape
if c != 3:
raise RuntimeError(f"{self} frame channels={c} do not match expected 3 channels (RGB/BGR).")
if h != self.capture_height or w != self.capture_width:
raise RuntimeError(
f"{self} frame width={w} or height={h} do not match configured width={self.capture_width} or height={self.capture_height}."
)
processed_image = image
if self.color_mode == ColorMode.BGR:
processed_image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
processed_image = cv2.rotate(processed_image, self.rotation)
return processed_image
def _read_loop(self):
"""
Internal loop run by the background thread for asynchronous reading.
On each iteration:
1. Reads a color frame with 500ms timeout
2. Stores result in latest_frame (thread-safe)
3. Sets new_frame_event to notify listeners
Stops on DeviceNotConnectedError, logs other errors and continues.
"""
while not self.stop_event.is_set():
try:
color_image = self.read(timeout_ms=500)
with self.frame_lock:
self.latest_frame = color_image
self.new_frame_event.set()
except DeviceNotConnectedError:
break
except Exception as e:
logger.warning(f"Error reading frame in background thread for {self}: {e}")
def _start_read_thread(self) -> None:
"""Starts or restarts the background read thread if it's not running."""
if self.thread is not None and self.thread.is_alive():
self.thread.join(timeout=0.1)
if self.stop_event is not None:
self.stop_event.set()
self.stop_event = Event()
self.thread = Thread(target=self._read_loop, args=(), name=f"{self}_read_loop")
self.thread.daemon = True
self.thread.start()
def _stop_read_thread(self):
"""Signals the background read thread to stop and waits for it to join."""
if self.stop_event is not None:
self.stop_event.set()
if self.thread is not None and self.thread.is_alive():
self.thread.join(timeout=2.0)
self.thread = None
self.stop_event = None
# NOTE(Steven): Missing implementation for depth for now
def async_read(self, timeout_ms: float = 200) -> np.ndarray:
"""
Reads the latest available frame data (color) asynchronously.
This method retrieves the most recent color frame captured by the background
read thread. It does not block waiting for the camera hardware directly,
but may wait up to timeout_ms for the background thread to provide a frame.
Args:
timeout_ms (float): Maximum time in milliseconds to wait for a frame
to become available. Defaults to 200ms (0.2 seconds).
Returns:
np.ndarray:
The latest captured frame data (color image), processed according to configuration.
Raises:
DeviceNotConnectedError: If the camera is not connected.
TimeoutError: If no frame data becomes available within the specified timeout.
RuntimeError: If the background thread died unexpectedly or another error occurs.
"""
if not self.is_connected:
raise DeviceNotConnectedError(f"{self} is not connected.")
if self.thread is None or not self.thread.is_alive():
self._start_read_thread()
if not self.new_frame_event.wait(timeout=timeout_ms / 1000.0):
thread_alive = self.thread is not None and self.thread.is_alive()
raise TimeoutError(
f"Timed out waiting for frame from camera {self} after {timeout_ms} ms. "
f"Read thread alive: {thread_alive}."
)
with self.frame_lock:
frame = self.latest_frame
self.new_frame_event.clear()
if frame is None:
raise RuntimeError(f"Internal error: Event set but no frame available for {self}.")
return frame
def disconnect(self):
"""
Disconnects from the camera, stops the pipeline, and cleans up resources.
Stops the background read thread (if running) and stops the RealSense pipeline.
Raises:
DeviceNotConnectedError: If the camera is already disconnected (pipeline not running).
"""
if not self.is_connected and self.thread is None:
raise DeviceNotConnectedError(
f"Attempted to disconnect {self}, but it appears already disconnected."
)
if self.thread is not None:
self._stop_read_thread()
if self.rs_pipeline is not None:
self.rs_pipeline.stop()
self.rs_pipeline = None
self.rs_profile = None
logger.info(f"{self} disconnected.")
@@ -0,0 +1,82 @@
# Copyright 2024 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.
from dataclasses import dataclass
from ..configs import CameraConfig, ColorMode, Cv2Rotation
@CameraConfig.register_subclass("intelrealsense")
@dataclass
class RealSenseCameraConfig(CameraConfig):
"""Configuration class for Intel RealSense cameras.
This class provides specialized configuration options for Intel RealSense cameras,
including support for depth sensing and device identification via serial number or name.
Example configurations for Intel RealSense D405:
```python
# Basic configurations
RealSenseCameraConfig("0123456789", 30, 1280, 720) # 1280x720 @ 30FPS
RealSenseCameraConfig("0123456789", 60, 640, 480) # 640x480 @ 60FPS
# Advanced configurations
RealSenseCameraConfig("0123456789", 30, 640, 480, use_depth=True) # With depth sensing
RealSenseCameraConfig("0123456789", 30, 640, 480, rotation=Cv2Rotation.ROTATE_90) # With 90° rotation
```
Attributes:
fps: Requested frames per second for the color stream.
width: Requested frame width in pixels for the color stream.
height: Requested frame height in pixels for the color stream.
serial_number_or_name: Unique serial number or human-readable name to identify the camera.
color_mode: Color mode for image output (RGB or BGR). Defaults to RGB.
use_depth: Whether to enable depth stream. Defaults to False.
rotation: Image rotation setting (0°, 90°, 180°, or 270°). Defaults to no rotation.
warmup_s: Time reading frames before returning from connect (in seconds)
Note:
- Either name or serial_number must be specified.
- Depth stream configuration (if enabled) will use the same FPS as the color stream.
- The actual resolution and FPS may be adjusted by the camera to the nearest supported mode.
- For `fps`, `width` and `height`, either all of them need to be set, or none of them.
"""
serial_number_or_name: str
color_mode: ColorMode = ColorMode.RGB
use_depth: bool = False
rotation: Cv2Rotation = Cv2Rotation.NO_ROTATION
warmup_s: int = 1
def __post_init__(self):
if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
raise ValueError(
f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."
)
if self.rotation not in (
Cv2Rotation.NO_ROTATION,
Cv2Rotation.ROTATE_90,
Cv2Rotation.ROTATE_180,
Cv2Rotation.ROTATE_270,
):
raise ValueError(
f"`rotation` is expected to be in {(Cv2Rotation.NO_ROTATION, Cv2Rotation.ROTATE_90, Cv2Rotation.ROTATE_180, Cv2Rotation.ROTATE_270)}, but {self.rotation} is provided."
)
values = (self.fps, self.width, self.height)
if any(v is not None for v in values) and any(v is None for v in values):
raise ValueError(
"For `fps`, `width` and `height`, either all of them need to be set, or none of them."
)
+65
View File
@@ -0,0 +1,65 @@
#!/usr/bin/env python
# Copyright 2024 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.
import platform
from pathlib import Path
from typing import TypeAlias
from .camera import Camera
from .configs import CameraConfig, Cv2Rotation
IndexOrPath: TypeAlias = int | Path
def make_cameras_from_configs(camera_configs: dict[str, CameraConfig]) -> dict[str, Camera]:
cameras = {}
for key, cfg in camera_configs.items():
if cfg.type == "opencv":
from .opencv import OpenCVCamera
cameras[key] = OpenCVCamera(cfg)
elif cfg.type == "intelrealsense":
from .realsense.camera_realsense import RealSenseCamera
cameras[key] = RealSenseCamera(cfg)
else:
raise ValueError(f"The motor type '{cfg.type}' is not valid.")
return cameras
def get_cv2_rotation(rotation: Cv2Rotation) -> int | None:
import cv2
if rotation == Cv2Rotation.ROTATE_90:
return cv2.ROTATE_90_CLOCKWISE
elif rotation == Cv2Rotation.ROTATE_180:
return cv2.ROTATE_180
elif rotation == Cv2Rotation.ROTATE_270:
return cv2.ROTATE_90_COUNTERCLOCKWISE
else:
return None
def get_cv2_backend() -> int:
import cv2
if platform.system() == "Windows":
return cv2.CAP_AVFOUNDATION
else:
return cv2.CAP_ANY