Kinematic calibration with the camera

The robot model puts each joint’s 0° where Elephant’s CAD puts it. On the real arm the zeros are a little different. A small error at a joint near the base moves the fingertips by several millimetres. This page measures the zero offsets of the lab arm with the lab camera.

Method

1. Register the camera on the six base plate screws and the six screws on top of the base (experiments/camera-calib/tool/track_screws.py). The camera then knows where the base is. It does not know the arm.
2. Move the arm through poses in free air (random joint angles, the gripper closed, at least 45 mm above the table, out of a keep-out box over the laptop). At each pose, take a snapshot and record the joint angles and the IMU. 39 poses in three batches, then 16 new poses to test.
3. Find the edges of the arm in each snapshot: subtract the background (the median of all the snapshots, because the arm moves and the scene does not), then Canny edges.
4. Draw the model’s outline at the recorded angles and keep the points that no other part hides (ray test). Fit the six zero offsets so that the outline points lie on the edges (distance map, clipped at 15 px, robust loss). Repeat with the new outline until it does not change.
5. Find the 9 holes in the gripper’s palm where the palm faces the camera (10 of the 55 snapshots): dark discs of the expected size, then the whole hole pattern matched at once, so that a hole cannot be confused with its neighbour (the holes are about 8 px apart). Fit the offsets to the outline and the holes together.

The tools are in experiments/kinematic-calib on the plush-pick branch: collect.jl (the poses), kc_model.py (the model, its outline and the camera), kc_holes.py (the palm holes) and kc_fit.py (the fit; --holes adds the palm holes).

Result (2026-10-09)

Joint	J1	J2	J3	J4	J5	J6
Zero offset (°)	+0.95	−1.54	+0.67	−0.43	+0.25	+1.63

The gripper body is not where Elephant’s model puts it. The camera puts it 2.4 mm nearer J6 along the J6 axis and 1.2 mm to the side. The palm holes and the screws on the back face of the body agree.

• Along the axis, it is the flange: the body’s back face sits on the flange face. A body 2.4 mm nearer J6 means a flange face 2.4 mm nearer J6: the URDF joint joint6output_to_joint6 is now 43.2 mm (Elephant: 45.6 mm). For the gripper and the fingertips this is the same as moving the body.

• To the side, it is the mount: joint6output_to_gripper_base is now at x −0.2, y 1.2 mm (z 34 mm as before). The gripper is held by LEGO-style connectors: it came off the flange on 2026-10-07 and was put back.

• Why the offsets alone did not fit: with the zero offsets only, the hole pattern had the right shape in each snapshot (1–3 px), but the whole pattern was 4–8 px off, and in a different direction for each pose. A shift of the gripper body removes this. A lens distortion term, a gravity sag of J2 and J3, or a tilt of the gripper do not.

• J1 is the clear result. All fits give +0.9 to +1.3°.

• J2 to J6 are weaker. A fit on three batches and a fit on all four differ by up to 0.4°. The arm is white on a white wall and a light laptop, so many outline points have no edge to match. Only the palm holes are sharp point features.

Test on the 16 poses of the fourth batch, with the fit on the first three:

	Palm holes (3 poses, 35 holes): error	Outline: mean distance to an edge
No offsets, Elephant’s mount	—	7.6 px
Outline fit only (the first result, below)	10.0 px	6.8 px
Outline + holes, gripper shifted	2.7 px (about 1.3 mm)	6.7 px

The outline score does not separate the models: about 6.7 px is the limit of the edge images. The palm holes do.

First result (outline fit only, the base plate screws only): +1.08, −0.20, −1.57, +0.34, −0.13, +0.64°. It moved the fingertips by 5 mm on average.

Where the offsets apply

• mycobot_description/config/mycobot_280_arduino/calibration.yaml (zero_offset) holds them.
• The Julia package applies them to every angle (encoder_error, JOINT_ZERO_OFFSET). The ATOM still gets its own angles, so the arm goes where the calibrated model says.
• The Control page draws the calibrated pose in the 3D view and the camera overlay when the lab service serves /lab/calibration.json. The numbers on the page and the goals it sends are the ATOM’s angles.
• The firmware does not apply them.

Not done yet

• The closed fingertips are 2 mm short in the model. From the body’s back face (the face on the flange) to the end of the plastic finger (the middle of the inner edge), closed: 115 mm measured (2026-10-08 and 2026-10-09), 113.0 mm in the model. Open: 95 mm measured, 95.2 mm in the model. Fully closed, the foam pads do not touch, as in the model (−0.70 rad: 6 mm between the plastic pad faces); the camera fit of the closed finger angle agrees (−0.693 rad).
• The J6 housing mesh was 6.5 mm short at the ATOM end and 3.7 mm long at the flange end. It is now corrected from caliper measurements (see Changes to Elephant’s model). The outline fits in this page used the old mesh. A fit that lets the J6 axis move from J5 does not improve the error on new snapshots (palm holes: 2.8 px, 3.2 px with the J6 axis free), so the kinematics stay.
• A gravity sag term for J2 and J3: the fits did not agree, so it is not in the model.
• The finger model (the J7 to finger angle map) and the pads that the user added.
• A camera fit and an arm fit together. The camera is fixed by the screws here.
• More snapshots with the palm toward the camera: only 10 of the 55 have it.