/pb29
physical result / 10 july 2026
Quiet Attractor
41 paths / 8.37 m pen travel
PB29 is a robotic painting-arm project built around a UFactory xArm 7. The immediate job is deceptively simple: see a sheet of paper, locate the current pen tip, find safe contact, and make a mark where the system expected it to land.
The long-term aim is a physical painting system whose internal state stays visible. Plans, uncertainty, selected primitives, camera evidence, and observed results should remain inspectable instead of disappearing inside an autonomous model.
It sits inside my wider painting practice, but this project is about the machine: the calibration stack, mark-making grammar, and feedback loop required before an AI can direct a brush in the physical world.
01 / the hard part
The physical truth layer
Generating a beautiful path is easy. Knowing where paper, tool, surface, and robot actually are is the hard part. A few millimetres of stale geometry can turn a drawing command into a broken nib or a collision.
PB29 keeps these truths separate because they go stale for different reasons. Moving a camera, changing a pen, retaping paper, or changing the mount each invalidates a different part of the chain.
paper
Where may it draw?
A visual mask becomes a separate, inset executable region.
tool
Where is the nib?
The current tip is calibrated relative to the robot flange.
surface
Where is contact?
Plane geometry guides planning; fresh visual evidence proves contact.
motion
Is the path safe?
Saved joint paths are validated, branch-locked, and replayed exactly.
02 / architecture
A glass-box painting system
Deterministic, testable primitives do the physical work. A language or vision model may eventually choose an operation—add a hatch, cool a region, try a glaze—but it does not improvise motor commands.
The observability layer is not only debugging. Showing what the arm sees, trusts, plans, and learns is part of the artistic premise.
03 / evidence
From rough tests to a registered drawing
Early hatch pages were messy but useful: they exposed stale contact assumptions, branch changes, dry nibs, and paths that looked valid in software but were wrong in the room.
Quiet Attractor is the first complete study carried through the current workflow: generated path, paper registration, validated motion, physical drawing, and an after-image with the system's coordinate evidence preserved.
04 / now
Working, open, later
working
Registered A4 pen drawing, fixed and wrist vision, current tool calibration, guarded paths, and a read-only operator console.
open
Automatic first contact for arbitrary tools and paper, a stronger surface model, and a useful catalogue of mark primitives.
later
Brush and acrylic handling, dip/wipe/rinse choreography, tool changing, force sensing, and high-level model direction.
context + traces