RealSense
Status: hardware-verified
Hardware-verified Color capture and recording are checked in real collection on Rakuda2.
RealSenseRGB uses pyrealsense2 to capture only the color stream. It requires the hardware extra.
type: realsense_rgb
serial: SET_REALSENSE_SERIAL
width: 640
height: 480
fps: 30
Replace serial with the serial number of the connected camera. If the SDK's rs-enumerate-devices is available, you can use it to check the number and connection information. Set width, height, and fps to a combination the device supports.
RGB8 images are captured and stored in observation.images.camera (when the node name is camera). A frame-drop warning is shown when connected over anything other than USB 3. This stream does not output depth images, point clouds, or camera calibration values.
The node also needs a max_age_ms entry in nodes/recorder.yaml; see "Recorder entry" in Sensors.
For connecting it to Rakuda2, see the robot page.
Image storage
Set storage: video | image in each image sensor's node YAML (RealSense, DIGIT, V4L2 including Pika wrist cameras). The default is video: the existing lossy MP4 encoding. image embeds lossless PNG frames in Parquet; use it for DIGIT/background pixel differences. Formats can be mixed in one dataset. Invalid values are refused at configuration load, including validate.
PNG generally needs more disk space and PNG compression/write work; actual size and throughput depend on image content and hardware. As an uncompressed RGB reference, 320×240×3 at 30 dataset fps is 6.9 MB/s per sensor (640×480: 27.6 MB/s); these are not measured PNG rates. LeRobot 0.7.0 uses PNG compression level 6 for image. Changing storage changes the feature dtype: choose a separate dataset output; resume into the previous schema is refused.
Live observations remain HWC uint8 RGB, and LeRobot dataset reads return CHW float32 in [0, 1] for either format by default. Checkpoint inputs remain VISUAL. Storage was tested with mock/synthetic frames only, without hardware.
The codec of video MP4s (H.264 by default, or AV1) is chosen with the recorder's video_codec.