roarm_draw

AI drawing robot: a Waveshare RoArm-M2-S (4-DOF ESP32 arm, ST3215 serial-bus servos) drawing with a felt pen on a vertical bamboo board, driven from a Jetson AGX Thor. Open-loop IK + direction-aware backlash compensation, closed-loop encoder verification, and wrist-cam visual servoing for mark registration.

Coordinate frame (vertical board)

+x = INTO the paper (pen depth), +y = viewer-LEFT, -y = right, +z = up. The drawing plane is y–z; x is the pen up/down axis.

CLI — arm <cmd>

~/bin/arm wraps arm.py. Motion commands drive the arm directly (retract-then-move safe travel); the rest subprocess to the scripts below.

cmd	what
arm home	center at safe clearance, pen up
arm status	print pose + servo loads (read-only)
arm center / arm park / arm retract	positioning
arm pen up|down	lift / lower pen
arm setup	hold at reference pose to set the 10 mm standoff
arm draw [img.png]	vectorize + draw an image (default: built-in test scene)
arm preview [img]	plan preview, no arm motion
arm square	compensated test square
arm star	servo-registered 5-point star
arm check	in-air accuracy / compensation verify
arm calibrate	wrist-cam servo calibration (Jacobian + contact pixel)
arm servo-demo	localize a dot from several views and re-ink it
arm stream	live global-cam MJPEG at http://192.168.1.40:8080/
arm photo [wrist|global]	grab one frame to the scratchpad

Layout

• roarm.py — RoArm control class: JSON-over-serial, bounds, move_verified (closed-loop), draw_to/draw_line (backlash-compensated, cartesian-subdivided), pen_up/pen_down, park.
• strokes.py — image → strokes pipeline: skeleton/contour/edge vectorize → order (2-opt, cut pen-up travel) → CanvasMap px→mm letterbox → execute.
• draw_image.py — Phase 1 driver (image → plan → draw, with y-mirror fix).
• servo.py / servo_draw.py — Phase 2 wrist-cam visual servoing (image Jacobian, mark localization, servo-registered star).
• calibration.json — measured source of truth: reference pose, standoff, depth, workspace, backlash constants, canvas bounds.
• servo_calib.json — wrist-cam Jacobian J, inverse, contact pixel p0.
• cam_stream.py, corner_teach.py, setup_standoff.py, plus characterization scripts (undertravel_char.py, compensated_verify.py, contact_visual.py, …).

Hardware notes

• Serial: /dev/ttyUSB0 @ 115200. Opening the port resets the ESP32 (arm may move).
• Cameras: wrist /dev/video0, global /dev/video2 (Innomaker U20CAM 1080p); heavy barrel distortion, purple cast (black ink reads purple). Wrist cam needs manual exposure (~200) to see ink on white paper.
• Under-travel is deterministic/repeatable/speed-independent → software-corrected via set_backlash() + draw_to(); residual closed via move_verified and vision.

Setup: square the board, run arm setup, dial the pen to a 10 mm standoff, go.