Workshop material for using coding agents to make small, understood and verified changes to robotics software, from a point-cloud bug to a bounded TurtleBot3 simulation.

_{Advanced Telerobotics Research Laboratory · Department of Computer Science · Kent State University}

The Advanced Telerobotics Research Laboratory prepared this material for a two-hour workshop at Kent State University on Nov. 7, 2026: a 45-minute talk with live demos, then a one-hour hands-on lab. The emphasis is on specifying, understanding, implementing and verifying software with an agent; ROS 2 supplies a concrete integration problem. Everything runs on a laptop with authored fixtures, an authored synthetic rosbag and a headless Gazebo simulation. No sensor or robot is needed, and none is operated.

Start here

Slides The 25-slide deck as a PDF, with the slide manuscript and talk outline.	Hands-on lab Six scripted steps, with the commands for macOS, Linux and Windows and the common fixes.	Participant workbook The point-cloud exercise, its contract and the explain-back questions.
ROS 2 demos The synthetic rosbag replay and the bounded TurtleBot3 application in headless Gazebo.	Run the workshop Facilitator guide, setup and a rehearsal plan.	Evidence What ran, where, with which versions, and what has not been tested.

Try it

The core exercise needs only Python 3.9 or newer. From this folder:

python3 -I exercises/check.py --solution
python3 -I exercises/check.py

The reference solution passes 12 of 12 checks; the starter fails two on purpose until you fix it. The lab runs the same exercise one step at a time, with a doctor check for every step:

bash lab/00-check.sh
bash lab/01-reproduce.sh

On Windows without Bash, use py -3 lab\lab.py 00 (any step number). Steps 03 and 04 run ROS 2 Jazzy and Gazebo in a container on any computer with Docker, x86-64 or arm64:

docker pull ghcr.io/atr-lab/agentic-workshop-station:2026.11
bash lab/03-replay.sh
bash lab/04-simulate.sh

To install Python, Git and Docker Engine and pull the image in one go, run bash scripts/install.sh on macOS or Linux, or scripts\install-windows.ps1 on Windows (command sheet). The image is public, so no sign-in is needed. Without access to the registry, build it from this folder with docker build -t ghcr.io/atr-lab/agentic-workshop-station:2026.11 -f lab/docker/Dockerfile . (about 10 minutes). The command sheet covers every operating system and the problems people hit most.

What's inside

Folder	Contents
docs/	Start page, slide manuscript, talk outline, workbook, facilitator guide, setup, rehearsal, prompt cards, answer key and verification
lab/	The six lab step scripts (run, doctor, help), standard-library Python, and the Docker station image
exercises/, solutions/	The deliberately incorrect starter, its checker and the reference solution
ros2/agentic_workshop	ROS 2 package: point-cloud diagnostic node, synthetic replay runner and the bounded TurtleBot3 controller, command guard and scenario checker
skills/	Four agent skills for Codex and Claude Code; .agents/skills and .claude/skills link to them
templates/	Interface contract and run report
evidence/	Saved reports and environment records from Nov. 3–5, 2026
scripts/	Install scripts for participant laptops (macOS, Linux, Windows), setup for an Ubuntu 24.04 ROS 2 station, and a fresh practice copy of this folder

Use it with an agent

Open this folder in your agent, read AGENTS.md, then ask it to use the relevant skill. An explicit path such as skills/unilidar-replay/SKILL.md works even when the client does not discover skills on its own.

Skill	Example request	Output
robotics-orient	"Use this skill to scope the UniLiDAR exercise in my ROS workspace."	Environment inventory, interface contract, acceptance checks
explain-robotics-pr	"Help me understand this PR before deciding whether it is ready."	Code-grounded explanation and unresolved questions
unilidar-replay	"Build and verify a diagnostic sector-clearance node from this recording."	Bounded implementation, reproducible replay, evidence report
turtlebot-capstone	"Implement the workshop's TurtleBot3 stop-before-obstacle application in simulation."	Complete application, launch and configuration, scenario checks, run report

What this does and does not show

The replay uses an authored synthetic PointCloud2 bag, not a capture from Unitree hardware, and the capstone runs a simulated TurtleBot3 Burger in headless Gazebo. Results establish the simulated configuration only. A physical trial needs a prepared area, a verified stop path and separate approval, and the agent stays outside the real-time control loop. See verification for exactly what ran.

License

Source code and scripts (lab/, exercises/, solutions/, ros2/, scripts/) are under the MIT License. Documentation, skills, templates, slides and evidence write-ups are under Creative Commons Attribution 4.0. The Advanced Telerobotics Research Laboratory and Kent State University names and logos are not covered by either license.

---

About the lab · Getting started in the ATR Lab · Development guidelines · All repositories

_{Advanced Telerobotics Research Laboratory · Department of Computer Science · Kent State University · www.atr.cs.kent.edu}