Teaching robots to say what they mean, and to know when they can't.

I build the layers that stop an unsafe plan before hardware moves.

Drag a goal. The robot replans.

Drag the start, a goal or the zone.

A watcher's guessGoal A 50%   Goal B 50%
Drag anything. It replans live.
0of 67

unsafe routes from language models got past the safety verifier.

0 of 599 safe routes refused
32of 32

certified legibility bounds hold when re-checked.

a ceiling, not a best guess
10of 10

sealed goals made the navigation shield halt instead of inventing a route.

0 unsafe recoveries sent on
Open to robotics and AI roles, and PhD positions

Two ways in.

For hiring teams

I'm hiring an engineer

Robotics and AI engineering in C++20, Python and ROS 2, with embedded firmware and PyTorch. Available on immediate notice, with UK, EU and US hours.

For research supervisors

I'm looking for a PhD student

MSc AI and Robotics with Commendation, University of Hertfordshire. Research on legible motion, verifying language-model planners, and human-robot interaction.

Every number here is regenerated from committed code and data. Where an earlier claim could not be checked, it was withdrawn.Read the corrections note
Selected work

Follow the route.
Three stops.

All projects and eight live demos
A scale model of a building floor. A dashed orange route leaves the map. A solid blue route runs from the robot to the goal.
Navigation safety

Detect, recover, or halt.

llm-nav-shield

A language model proposes a route. The shield checks it, replaces it when it is wrong, and stops when no safe route exists.

38 / 40bad routes replaced with a verified one. None unsafe sent on.
A robot on a dark floor curving around a wall towards one of two goal rings, with fainter candidate paths beside it.
Legible motion

How clear can a robot's motion be?

legibility-bounds

Certified upper and lower bounds on legibility under a path budget, so a hard world can be told apart from a weak planner.

32 / 32bounds hold. The gap between them is reported, not hidden.
A scale model of a house. A robot follows a blue route through the living room to a door marked with an orange bar.
LLM planning

How do language-model planners fail?

plan-failure-bench

One planted trap per instruction, answers in a checkable action language, and every label a proof. No human or model does the judging.

6trap families under one protocol, each with its false positive count.
Who you would be working with
Portrait of Munawar Kazmi

I work at the boundary between robotics research and shipped engineering.

“From the initial stages of designing interactive scenarios for the robot to conducting experiments with human participants and analysing results, Munawar demonstrated remarkable technical proficiency, creativity, and a strong work ethic.”Dr Patrick Holthaus Reader in Interactive Assistive Technology, University of Hertfordshire. MSc supervisor.

About me    All references

Goal reached

Have a robot that needs
to know its limits?

hello@munawarkazmi.com
GitHubLinkedInORCIDMultan, Pakistan. Calls in UK, EU and US hours.