SanMower started with a practical question: could a conventional 21-inch electric mower deck become a useful autonomous platform without turning the project into a large robotics framework? The resulting design uses a Greenworks brushless deck, differential-drive hub motors, a Raspberry Pi 3 for autonomy, and a Raspberry Pi Pico for the control loop that must remain predictable even when the Linux computer is busy.

Splitting the system by responsibility

The Raspberry Pi is the high-level computer. It reads GPS data, reads the BNO055 orientation sensor over I2C, maintains the mower’s position and heading estimate, generates lawnmower-pattern waypoints, and applies the virtual geofence. It also provides the communications layer for MQTT telemetry and the optional low-overhead Flask/HTMX monitoring page.

The Pico is deliberately narrower in scope. It receives commands from the Pi over UART, produces PWM and direction signals for the left and right drive controllers, controls the blade controller, and watches the physical inputs that should not depend on a general-purpose operating system. Those inputs include three normally closed bumper or limit switches, the auto/manual selector, and the emergency-stop circuit.

This boundary makes the system easier to reason about: the Pi can decide where to go, while the Pico is responsible for making motor outputs and stop conditions behave consistently.

Power and hardware layout

The design uses a 56V EGO battery bus. The blade controller is connected to the high-voltage bus, while a 56V-to-36V converter supplies the two RioRand drive controllers. A separate converter provides the 5V logic rail for the Pi, Pico, and receiver. The two hoverboard hub motors provide differential drive, and the mower’s brushless motor remains the cutting actuator.

The FlySky receiver supports a useful staged test workflow:

  • manual driving through the receiver while validating steering and direction;
  • recording a perimeter from GPS points;
  • autonomous waypoint following after the boundary and interlocks have been checked.

Manual drive and blade output are intentionally limited during this phase. The receiver’s emergency channel is also treated as a stop input rather than merely another user-interface command.

The mower combines GPS position with IMU orientation so it can maintain a more useful heading estimate during turns and short periods of noisy GPS data. Boundary profiles and geofence polygons are intended to remain in memory, keeping containment checks quick and avoiding unnecessary SD-card writes on the Pi.

The monitoring surface is intentionally small. Flask serves the page and status endpoints, while HTMX polls current state without a front-end build pipeline. The page can show mowing state, coverage progress, and a generated geofence/track map, and can send pause or emergency-stop commands over the local network. MQTT remains available for machine-to-machine telemetry and future dashboards.

Lessons from the wiring review

The most important lesson is that software safety logic is not a substitute for electrical safety engineering. The project documentation identifies several items that must be resolved and tested before operating a blade:

  • add a safety-rated contactor or equivalent hardware path that can remove motor power;
  • add dedicated branch protection for the blade and drive circuits;
  • verify that the RioRand speed inputs and VESC PPM input accept the Pico’s 3.3V signals;
  • validate current, peak load, thermal, wire-gauge, grounding, and pre-charge assumptions;
  • define and test receiver-link-loss behavior.

These checks are part of the project, not afterthoughts. The goal is a mower that can be observed and stopped predictably before autonomy is treated as successful.

SanMower is therefore both a robotics build and a systems-engineering exercise: mechanical constraints, power electronics, embedded firmware, navigation, telemetry, and operator controls all have to agree on what “safe to move” means.