Best camera for robot learning
For imitation learning with LeRobot or ALOHA-style setups, a plain USB webcam is the right default: the Logitech C922 is what ALOHA and Mobile ALOHA use, and a small wide-angle USB module makes a good wrist camera. Add a depth camera such as the OAK-D-Lite or a RealSense only if your policy or navigation actually uses depth, since most popular imitation learning policies train on RGB alone.
What matters
- Software support: LeRobot reads any UVC webcam through OpenCV and Intel RealSense through its own class; other depth cameras need extra code.
- Fixed exposure and steady frame rate matter more than resolution; policies usually train on small images.
- Wrist cameras need a wide field of view, low weight and a short cable path.
- USB bandwidth: three or four cameras on one USB controller can drop frames, so use separate ports or a powered hub.
- Depth only if you use it: depth cameras cost 2 to 4 times more than webcams.
Our picks
| Pick | Why | Where to buy |
|---|---|---|
| Logitech C922 / C922x Pro Stream Top and front views on a fixed workcell | The camera ALOHA and Mobile ALOHA were built and tested with, so recipes and datasets match. Plain UVC, so it works in LeRobot with no drivers. | Logitech: $119.99 Amazon: check price |
| Arducam 1080p wide-angle USB camera (IMX291) Wrist cameras | Small board camera with a very wide lens that sees the gripper and the object together; ToddlerBot uses a 160 degree IMX291 model. | The Pi Hut: GBP 60.60 (unverified) Amazon: check price |
| Luxonis OAK-D-Lite Depth plus onboard AI on a mobile robot | Stereo depth and a 4K RGB sensor with its own neural accelerator, so a Raspberry Pi can run detection without a GPU. Standard on the TurtleBot 4 Lite. | Luxonis: $269.00 |
| Intel RealSense D405 Close-range depth on a wrist | Tuned for 7 cm to 50 cm, which suits a gripper view, and supported directly by LeRobot's RealSense camera class. It has no separate RGB module, so manual color controls do not work. | RobotShop: check price (unverified) no current US price |
| Intel RealSense D415 Head or room-scale depth | Longer-range depth with a rolling-shutter RGB camera; XLeRobot uses one as its head camera. | CDW: $381.99 (unverified) |
FAQ
Do I need a depth camera for imitation learning?
Usually not. LeRobot's own SO-101 and LeKiwi setups use ordinary RGB webcams. Depth helps for navigation, grasp planning, and policies built to take depth as input.
How many cameras should I use?
Two is a common start: one fixed overhead or front view for context and one on the wrist for the grasp. ALOHA uses four. Keep fixed cameras in exactly the same place between recording and testing.
Does a RealSense work on a Mac?
LeRobot's docs say RealSense on macOS is unstable and may need sudo to detect the camera; Linux is the safer choice.
What resolution and frame rate should I record?
LeRobot's camera examples use 640x480 at 15 to 30 fps for depth cameras. Higher resolution mostly costs bandwidth and storage, since policies downscale images anyway.
Sources: huggingface.co · huggingface.co · github.com · www.logitech.com · shop.luxonis.com · github.com · eu.robotshop.com · thepihut.com