Deskinator

Deskinator was a self-directed robotics build around a deceptively strict brief: create a compact, touch-free vacuum robot that can start with a hand gesture, discover an unknown rectangular tabletop, clean it without crossing an edge, and aim for a two-minute cycle. I designed and implemented the complete system—from electronics and mechanical integration through the Raspberry Pi software, simulator, analysis, and final documentation.

Rather than assuming a known map, the robot earns one. It follows the boundary, records where its two front sensors lose the surface, fits a rectangle, contracts that rectangle using the physical sensor-to-vacuum geometry, and only then executes a coverage path. Safety is therefore a property of perception and planning, not a last-second motor stop.

The supported new-alg executable is deliberately organized around one synchronous 50 Hz loop in DeskinatorSimple.run(). Every cycle checks the gesture stop, reads edge sensors, advances differential-drive odometry from step counts, evaluates the state machine, commands the motors, rasterizes the vacuum footprint, and writes telemetry. A shared time base makes a failure traceable from sensor sample to actuator command.

The executable states are explicit: WAIT_START, BOUNDARY_DISCOVERY, COVERAGE, and DONE. A failed rectangle fit enters DONE; a stop gesture exits the active loop into the same safe-shutdown path. Optional IMU and EKF files remain experimental scaffolding; the supported default path uses stepper odometry.

Two front APDS9960 sensors report whether the surface is present. The filter combines a raw threshold, an exponentially weighted moving average, hysteresis, and 60 ms debounce before an observation can become an edge event. Each event is projected into world coordinates from the robot pose and measured sensor offsets, preserving which physical reading produced each boundary point.

After the robot closes a lap, SimpleRectangleFit extracts candidate lines with a 2 cm RANSAC inlier threshold, refines each direction with SVD/PCA, searches for orthogonal four-line combinations, and scores how well they explain the observations. A coarse 5° angle search provides a fallback when the line combination is weak.

CoveragePlanner derives a safe inset from the sensor-to-vacuum offset, then lays lanes along the rectangle’s longer dimension. A 0.20 m cleaning width with 0.02 m overlap controls the lane spacing. The planner evaluates four traversal variants to reduce the distance from the current pose, alternates lane direction, adds explicit turn transitions, and finishes with a perimeter pass.

The drive-then-turn controller turns those oriented waypoints into linear and angular velocity commands. StepperDrive converts them to wheel rates, applies acceleration-aware updates, and emits 1/16-microstep pulses through A4988 drivers. In parallel, a 5 mm SweptMap raster and CSV telemetry expose where the modeled vacuum actually traveled.

The reported 50-run simulation campaign retained 48 completed trials. On an ideal 2 m × 2 m table with randomized starting poses, those runs averaged 99.40% coverage of the safe inset and 90.97% of the full modeled tabletop. Mean completion time was 137.22 seconds, so the planner missed the two-minute target even while producing strong simulated coverage.

• The simulator uses deterministic unicycle dynamics and does not model sensor noise, wheel slip, battery sag, or surface variation; these are simulation results, not physical-cleaning claims.
• Rectangle fitting and coverage planning assume a rectangular tabletop.
• The source workbook’s stale summary and invalid rectangle-error field are excluded; the published statistics come from the 48 retained raw rows.
• Physical performance remains sensitive to surface reflectivity, wheel traction, sensor placement, and battery state.