Dhruv Mehta, PhD

I like problems that only count as solved when a real machine moves. That has taken me across most of the modern robotics stack rather than down one lane of it. My PhD was reinforcement learning for ground vehicles — hybrid learning for rough terrain, a single agent coordinating two robots, terrain I authored myself so the training was worth trusting. Since then the interesting questions have moved up the stack: vision transformers and depth networks fast enough to run on a Jetson while everything else is also running, and vision-language-action models — OpenVLA, OpenPI — that turn an instruction into motion instead of a hand-written waypoint list.

What ties it together is that I want to be responsible for the whole path, from the loss curve to the robot in the room. A policy that only works in the simulator isn’t finished, and a demo that works once isn’t either.

The reason I stayed in this field is less academic than that. Physical AI is finally at the point where robots can take on work that is dull, dirty or genuinely dangerous for people — off-road logistics, last-mile delivery, inspection in places nobody should have to stand. I want the systems I build to make someone’s day materially easier, and I’d rather ship something reliable and modest than demo something spectacular that only works once.

I also like the parts of this job that aren’t code. I’ve taught two graduate robotics courses at Clemson, co-chaired student activities for IEEE RAS across ICRA and IROS, worked directly with NVIDIA’s robotics team on an integration, and demoed autonomous vehicles to the public at Innoventure and Auto Tech Day. Explaining a system to someone who doesn’t share your vocabulary is its own engineering skill, and it’s one I’ve deliberately practiced.

If you’re working on something in this space — hiring, collaborating, or just stuck on a problem that won’t behave — say hello. I’m always happy to talk shop.