Building robots to make the world better.
I am a research scientist at Nuro, on the Behavior Planning team, post-training flow-based driving policies with reinforcement learning. Before that, an MS at the CMU Robotics Institute and a dual degree at IIT Kharagpur. Along the way I have been lucky to work with Sanjiban Choudhury, Red Whittaker and Andrea Bajcsy.
I am optimistic about a future where robots will fundamentally change the way humans live, the way the industrial revolution made modern cities possible. But to get here, we need robot policies that self-correct and improve from the mistakes they make when executing tasks. Humans already do this: we run a locally optimal day-to-day policy that needs no deliberation, and a slower kind of reasoning that revisits what went wrong and sharpens the fast one. I think of this as Acting Fast and Slow, and I believe it is the fundamental nature of robotics policies that will actually generalize to produce the GPT moment for robotics.
I have tried to explore various versions of the idea in the past, with a few projects at Nuro, at hackathons and also my research on planning to recover from mistakes. As we mature towards bigger and better models, I really like the current shift towards world action models, which arrive at the same structure from the other side: a model that can imagine consequences is exactly the slow half a fast policy needs to check itself against.
PSI like building stuff on the side, and keep experimenting with fun ideas on my SO101 arm. Hit me up if you have any cool ideas I can try out.
Mahindra e2o2018 to 2020
Robotaxi footage from Nuro, Robotaxi Engineering Vehicle, Mountain View, CA (20x speed). Used with credit; video belongs to Nuro.
Robust imitation via learning to search
A learning-to-search method: at every control step the robot searches over imagined latent rollouts for a correction to its own plan, so it can notice it is going wrong mid-episode and steer back. It beats diffusion policies trained on 5 to 10x more data.
Autonomous regolith excavation on lunar-like terrain
An autonomous excavator that builds berms out of lunar regolith simulant: a bucket drum that shaves material continuously to hold ground reaction forces low, and a planner that decides where to dig and in what order.
ACTing fast and slow, hierarchical LLM-VLAs for long-horizon tasks
A two-level system: short-horizon visuomotor specialists trained with ACT, and an LLM above them deciding which one to run next and when it has failed. Runners-up at the Physical AI Hack.
All-purpose terrace farming robot
A terrace farming robot that lifts itself up 40 cm steps on a pair of scissor lifts, then plows, sows, rolls, waters and sprays the terrace it just climbed onto.
Self-fabricated autonomous ground vehicle
A ground vehicle built from bare frame to full autonomy stack by a student team, and runner-up at the Intelligent Ground Vehicle Challenge in Michigan two years running.
