Robotic Art VLA Group — BLUEBOT Platform

Overview

Painting is a useful benchmark for robot learning: the task is continuous, contact-rich, visually self-evaluating, and tolerant of stylistic error. Our group studies these properties on BLUEBOT, a SCARA-configuration industrial arm built around April 1983 and donated to the lab as non-functional surplus. Over roughly two months the machine was reverse engineered, refitted with modern control electronics and servo loops, and integrated with a LinuxCNC motion stack.

The resulting platform gives us a rigid, repeatable, human-scale canvas manipulator for a fraction of the cost of a modern collaborative arm — and a hardware substrate transparent enough that every layer, from encoder counts to stroke policy, is instrumentable.

BLUEBOT SCARA arm painting on canvas — **Figure 1.** The restored manipulator executing an acrylic stroke sequence on stretched canvas. Tool-center-point repeatability is sufficient for multi-layer registration without fiducial re-referencing.

Experimental Platform

The arm is a three-theta SCARA: each rotary joint carries a brush-DC servo motor and a 500 count per revolution incremental encoder. No Z axis was present in the original configuration, so a vertical stage was designed and fabricated around a recovered Parker 402006 linear stage to provide brush lift and pressure authority.

Parameter	Value	Note
Kinematics	SCARA, 3× revolute (θ₁–θ₃) + prismatic Z	Z axis added in-lab
Vintage	Dainichi Kiko PT-300H, ca. April 1983	Distributed stateside via PAR Systems
Actuation	Brush-DC servos; NEMA 23 retrofit path	3D-printed motor adapter
Feedback	500 count/rev incremental encoders	Quadrature, per-axis
Motion control	LinuxCNC + custom servo loops	Step/dir and analog velocity modes
Transmission	Harmonic reduction gearboxes	One unit rebuilt after wear failure

Robot arm during electrical revival and bring-up — **Figure 2a.** Bring-up: original 1980s drive cabinet bypassed in favor of an instrumented, externally commanded servo architecture.

CAD model of DC servo to NEMA 23 motor adapter — **Figure 2b.** Printed adapter mapping the legacy DC servo footprint to a NEMA 23 bolt pattern, enabling drop-in modern actuators.

Harmonic gearbox disassembled for replacement — **Figure 2c.** Harmonic drive teardown. Backlash growth in this stage set the practical floor on our stroke-placement error budget.

Control electronics and drive enclosure — **Figure 2d.** Amplifier and interface electronics packaged clear of the work envelope to keep the canvas region unobstructed.

Axis limit and homing hardware — **Figure 2e.** Homing and limit hardware. Absolute re-referencing at power-on is a prerequisite for comparing learned policies across sessions.

Scanned 1984 manufacturer documentation for the PT-300H — **Figure 2f.** Recovered 1984 manufacturer documentation used to reconstruct joint limits, gear ratios, and wiring conventions.

Research Method

1 · Stroke primitive library

Human painters are recorded producing isolated strokes across brush types, loading conditions, and approach angles. Each stroke is parameterized by entry pose, pressure profile, velocity envelope, and release, forming a discrete primitive vocabulary the planner can sequence.

2 · Learning from visual outcome

An overhead camera closes the loop on the canvas rather than the joint space. The learner scores its own output against a target rendering, then updates the mapping from intended mark to commanded trajectory — absorbing brush compliance, paint rheology, and drying behavior that are impractical to model analytically.

3 · Hardware-in-the-loop constraint

Because the manipulator is a rigid legacy machine, its dynamic limits are fixed and well characterized. This makes it a useful adversary: policies must respect real acceleration ceilings and residual backlash, not idealized simulator dynamics.

Design position. Reclaiming decommissioned industrial hardware lowers the capital barrier to embodied-learning research. A donated 1983 arm plus open-source motion control reaches useful precision for expressive-motion studies at a small fraction of new-equipment cost.

End effector brush holder over palette and canvas — **Figure 3.** End-effector brush mount, palette, and canvas fixture. Palette geometry is registered in the same frame as the canvas so paint pickup is a planned motion, not a manual step.

Results

Full closed-loop operation of all three theta axes restored, with a fabricated Z stage providing brush-pressure control.
Multi-layer acrylic paintings completed autonomously from a target image, including palette pickup, stroke sequencing, and layer ordering.
Platform submitted to the RobotArt painting competition in 2017 and 2018, providing an external, juried evaluation of output quality.
Failure analysis of the harmonic transmission established the dominant error term in stroke placement, guiding the current rebuild.

Completed robot painting on canvas — **Figure 4a.** Completed autonomous acrylic work.

Second generation upgrades to the arm — **Figure 4b.** Second-generation mechanical and electrical upgrades.

Motor and printed adapter installed on the arm — **Figure 4c.** Retrofit actuator installed on the shoulder axis.

Media

Recorded runs of the platform painting under autonomous control.

Clip 1. Full autonomous painting run, RobotArt 2017 submission (source).

Animated CAD spin of the motor adapter — **Clip 2.** Animated CAD of the servo-to-NEMA 23 adapter used in the actuator retrofit.

Data, CAD & Code

Mechanical CAD — servo-to-NEMA 23 adapter, Z-axis mount, brush holder. STL / STEP
Motion configuration — LinuxCNC .ini / .hal files, per-axis servo tuning.
Stroke dataset — demonstrated stroke primitives with synchronized joint and camera logs.
Reference documentation — reconstructed PT-300H kinematic and electrical notes.

Primary build log and downloads: transistor-man.com/bluebot_revival.html

Selected Writing & Reports

Reviving an Ancient Robot for Art: Reverse Engineering a 1983 SCARA Manipulator

Technical build log, 2017. Full write-up
Harmonic Drive Wear as the Limiting Term in Robotic Stroke Placement

Internal report, Robotic Art VLA Group.
Retrofitting Legacy Brush-DC Servo Axes for Open-Source Motion Control

Methods note, with CAD and configuration release.

People & Acknowledgements

Dane Kouttron — Program Engineer, principal investigator on the BLUEBOT platform.

The manipulator was donated to the lab by MIT alumni, and its recovery, transport, and bring-up depended on volunteer collaborators who moved several hundred pounds of 1980s robotics out of a basement and into a working shop. We thank them.

Learning Brushstroke Control on a Reclaimed Industrial SCARA