Wrist Camera Setup (SO-101 / LeRobot)

Camera configuration for imitation-learning robotic arms, identified as the #1 success factor separating a working policy from a failing one on so-101 in lerobot.

The two-camera standard

• Front (3rd-person): a fixed camera viewing the entire scene — the arm plus the workspace.
• Wrist (first-person): a camera mounted on the wrist that sees what the gripper sees.

A single fixed camera causes the arm to "lose sight" of the object as it approaches to grasp, making the policy hesitate. The wrist cam provides visual feedback up to the last millisecond of the grasp.

Practical hardware options

Option	Description	Cost
A	Laptop webcam as front + existing USB cam on wrist (prop laptop on books at ~45°)	0€
B	Cheap 2nd USB webcam (e.g. Logitech C270) on tripod as front	~20€
C	Dedicated UVC 32×32mm 720p/30fps module — format all official mounts support	~15€

3D-printable wrist mounts

The Robot Studio publishes 3 official wrist-mount variants in the SO-ARM100 repo, including the Hex-Nut Recess mount (SO-101 only, tool-free disassembly), designed around the UVC 32×32mm module.

Relation to transfer

Consistent camera placement between source and target embodiments improves zero-shot and fine-tuned transfer quality. See cross-embodiment-learning.

See also

• so-101
• lerobot
• cross-embodiment-learning
• act-policy
• imitation-learning