V1 · Built and driven
The robot I built
A welded stainless-steel frame on four mecanum wheels, so it drives in any direction and turns on the spot. Two computers share the work: an STM32 motor board runs everything with a deadline, and a Raspberry Pi 5 runs everything that needs a map, a model or a network.
Firmware
My own firmware for the STM32 motor board, in place of the vendor's closed one: motor PID at 100 Hz, quadrature encoders, the IMU, wheel odometry, a binary serial protocol, settings kept in flash, and a 2-second timeout that stops the motors if the host goes quiet.
12 ROS 2 packages
An EKF that fuses wheel odometry with a Madgwick-filtered IMU, mapping with slam_toolbox and AMCL on saved maps, Nav2 with the MPPI controller in its omnidirectional mode, and a frontier explorer re-implemented in Python from m-explore-ros2.
The dashboard
React and TypeScript, reaching ROS 2 through rosbridge over a WebSocket, with the camera over WebRTC: battery, motors, odometry, the IMU, the live map, the camera, and driving. Every command goes through the same ROS 2 topics as everything else.
orbibot_agent
An AI layer with 58 robot tools that runs on Gemini, Claude, Kimi or a local model through Ollama, with three ways in: an HTTP API, a terminal client and an MCP server, and a safety layer on its direct motion tools.
Short silent clips from the build log. Tap one to play it.
Measured on the bench, with the wheels off the ground, as the report describes.
Put the deadlines on the microcontroller
PID, encoders and the stop timeout run on the STM32. By design, a busy Pi or dashboard shouldn't make the wheels stutter (not yet tested under load), and the next robot's computer can change without touching the firmware.
A 2D LiDAR sees one flat slice
A table top above it or a low box below it does not exist to the LiDAR. That is why V1 got a depth camera.
Maps in a browser are a bandwidth problem
The dashboard was designed around what the robot has to serialise and send, not around how the page looks.