On the real arm

The skill, the key, and the YAML do not change. Two measurements do, because a policy trained on a real wrist camera is aimed using a lens model and a camera-to-fingertip transform that simulation invented for you with CameraMount. Hardware has to measure them.

You need the arm from Tutorial 3, the inverse kinematics from Move to a pose, and ik_tcp set to UmiSlimFeetechModule.tcp_from_flange. Hand-eye solves for the fingertips. Without that offset it solves for the flange and every reach is short by the length of the gripper.

The lens

Stop run-robot. The command opens the CSI camera itself, and the robot would already be holding it.

solarian --deployment lab calibrate-umi-camera

It walks you through a printed checkerboard, fits a fisheye model, and writes a remap file that the camera YAML then points at. The remap is what makes this sensor's image look like the GoPro image the policy was trained on. A policy shown raw IMX477 frames will reach for an object that is not where the pixels say it is.

The prompts are the procedure. Print the board at 100 percent, not fit-to-page. When the command finishes, conf/modules/right_wrist_camera.yaml should reference the new remap. Do not delete that block.

Hand-eye

Hand-eye is where the fingertips are in the camera's frame. The policy treats that transform as a constant baked into every training episode. If yours differs, the hand converges on a point next to the object and stays there, which looks like a stubborn model.

Print UMI's table marker at 100 percent. The default square is 160 mm on a side; measure yours and pass marker=<millimetres> if it is not. Tape it flat where the arm can see it from several angles, and do not move it during the solve.

Stop run-robot again. This command moves the arm:

solarian --deployment lab calibrate-hand-eye

You set one pose by hand so the camera sees the marker. The arm then orbits the marker through a spread of tilts and rolls. It writes conf/calibrations/right_wrist_camera_hand_eye.json and prints what to do with the result.

Read that printout. It compares the measured fingertip position, in the camera frame, with the position a UMI policy assumes. A difference of centimetres is the mount or the marker size: the policy commands deltas in tool axes, and no YAML key soaks up a camera that is bolted in the wrong place. If the printout tells you to change ik_tcp, put that orientation on RightArmModule(ik_tcp=...) and restart the robot. If it says the rotation is inside the noise, leave ik_tcp as the gripper's tcp_from_flange.

Re-run calibrate-hand-eye after any change to the remap. The axes move with it.

Details and the flags (mode=manual if you would rather pose every sample yourself) are in the command's --help. The skill does not open the JSON itself. It commands the TCP the arm publishes, which is the flange plus ik_tcp. The JSON is the record of the solve, and the printout is the part you act on.

The same key

Put the object the checkpoint was trained to pick in front of the gripper, in view of the wrist camera. On the Pi, run the relay, the robot, and the policy server. The server has to be able to read the checkpoint path in lab.yaml. A path that exists only on your laptop will not.

solarian --deployment lab run-server
solarian --deployment lab run-robot sim=false
solarian --deployment lab run-service pick

lab.yaml needs the same two keys as localhost, with values that make sense on the Pi:

policy_server_url: http://127.0.0.1:8001
policy_checkpoint: /path/on/the/pi/pick.ckpt

The skill file stays conf/skills/pick.yaml. It already interpolates those keys. If the server runs on the Pi and the controller runs on the laptop, policy_server_url in the laptop's deployment must be the Pi's address, not 127.0.0.1. 127.0.0.1 is correct only for the machine running run-service.

On the laptop:

solarian --deployment lab run-controller

Press 4.

The arm should reach for the object and close. Press 0 for the joint home from MoveSkill if you want it to let go of the behavior and park. Keep a hand near Control-C on the robot terminal for the first few runs. A policy that was going to miss will still move at max_speed.

If the hand consistently stops short or to the side, redo hand-eye before you retrain anything. The checkpoint did not change between simulation and this chapter. The camera did.