software engineer - robot builder

Daniel Osman · LinkedIn

domains

projects

Five-minute talk for Śniadania Robotyczne. Covers reflexes, perception and reasoning with examples from my four robot builds.

ai: local llms · vision-language models

Another Simple State Manager: a TypeScript library for shared and local state, with immutable snapshots, filtered subscriptions and microtask-batched notifications.

software: typescript · state management

Experimental TypeScript city simulator with a browser dashboard, inspired by Donella Meadows' book Thinking in Systems. Resources flow through reinforcing and balancing feedback loops.

software: typescript · web apps · data visualization · simulation

github.com/danielosman/city-sim ↗

Invoice processing with local LLMs: PDF to Markdown to schema-validated JSON, then reasoning over the typed data.

software: typescript · web apps
ai: local llms · document processing

Mecanum-wheel rover on a Raspberry Pi Pico 2 W with a 1D LiDAR. No gyroscope or motor encoders, yet orients itself in space by correlating current and past room distance traces.

electronics: raspberry pi pico
sensors: lidar
control & navigation: mapping · mecanum wheels
software: typescript · web apps · data visualization

Two-wheeled balancing robot on an Arduino Mega, with a gyro, stepper motors and a PID controller.

electronics: arduino
sensors: gyroscope
control & navigation: pid control

Arduino Uno and three analog distance sensors. If any sensor reads under 50 cm, it stops and turns towards the one with the largest reading. Fully deterministic code, yet its path through the flat was never the same twice.

electronics: arduino
sensors: analog distance sensors
control & navigation: obstacle avoidance

keywords

all projects

RoboCar 3

Map

The 8 × 8 distance sensor builds a map of 10 cm occupancy cells in three height layers.

3D view of a recorded drive: coloured distance points and orange blocks for occupied map cells — 3D replay of a recorded drive: distance readings as points, occupied cells as blocks.

Control

Web app: click a destination on a flattened 2D view of the map.
Command-line interface: send steering commands and read their outcomes. An AI agent can steer the robot the same way.

The web app: a 2D map of coloured cells, the camera picture and an 8 × 8 grid of distance readings — Web app: 2D map with detected objects as coloured rectangles, camera snapshot and 8 × 8 distance readings.

Movement tracking

In this mode the robot rotates to follow movement detected by the distance sensor.

Object detection

A local LLM finds and names objects in the camera picture and draws a bounding box around each. The box gives an estimate of where the object is on the 2D map.

In progress: giving the LLM tools to correct the object's position and orientation on the map.

Camera picture with a box around a sofa, labelled sofa 95% 365 cm — A bounding box from the local LLM.

2D cell map with the sofa marked as a box near its edge — A detected sofa placed on the 2D map.

Safety stops

The motors switch off when:

PicoB gets no heartbeat from PicoA
a wheel stops turning
the robot tilts more than 15°
the robot would drive into an occupied cell remembered on the map