AI Robotics Command Platform
For a university research lab, COET built the command layer between people and a Unitree Go2 quadruped: an operator interface that holds the platform's documentation and directs the robot in natural language, typed or spoken, with an agentic AI deciding how to act on what the robot senses.
Product context
Natural language in, robot behaviour out.
Researchers wanted to give the robot instructions the way they'd ask a person, and have it use what it can see. That meant a command path from speech or text, through an AI that can reason about the robot's sensors, down to the robot itself.
COET's contribution
Web interface
One place for the project's documentation and for talking to the robot by typing.
Embedded voice agent
A voice agent in the interface listens to spoken commands and passes them on.
Local AI inference
An Ollama server on the lab's own hardware runs the models the agent reasons with.
Agentic decision-making
An agent harness interprets each request, decides what to do, and turns it into robot commands.
Multimodal perception
The robot's LiDAR and camera feed the decisions, so it can act on what it sees.
How it works
Speak or type; the agent decides; the robot moves.
A command from the web interface or the voice agent reaches the agent harness. It reasons with a model served locally by Ollama, using the robot's LiDAR and camera, then sends commands down to the Go2.
Key features
- 01
Two ways to talk
Type in the web interface or speak to the voice agent.
- 02
Documentation built in
The project's docs live in the same interface.
- 03
Local inference
Models served by Ollama on the lab's own machine.
- 04
Agent harness
Interprets requests and decides on actions.
- 05
LiDAR and camera
Perception that feeds every decision.
- 06
Multimodal
Language and vision in one loop.
