People · Co-founder & CTO · Simulation & learning
Matt Bennice.
Simulation-driven robot control, imitation and reinforcement learning at fleet scale, and the sim-to-real discipline behind robots that work outside the lab. Google X, Everyday Robots, DeepMind, Wing, Stack AV, 1X, Cruise.
Simulation-driven controlImitation learningReinforcement learning at scaleSim-to-real transferFleet learningHuman-robot interaction
01 / Selected publications
The papers.
Learning to Learn Faster from Human Feedback with Language Model Predictive Control
Deep RL at Scale: Sorting Waste in Office Buildings with a Fleet of Mobile Manipulators
Jump-Start Reinforcement Learning
Practical Visual Deep Imitation Learning via Task-Level Domain Consistency
VADER: Visual Affordance Detection and Error Recovery for Multi Robot Human Collaboration
Interactive Multi-Robot Flocking with Gesture Responsiveness and Musical Accompaniment
Full record on dblp ↗
02 / Granted patents
The filings.
Simulation driven robotic control of real robot(s)
Simulation driven robotic control · continuation, granted 2025 under Google DeepMind
Re-simulation of recorded episodes
Imitation learning for training and refining robot control policies
Co-inventors include Chelsea Finn, Sergey Levine, Eric Jang and 7 others.
Simulating multiple robots in virtual environments
Injecting noise into robot simulation
Single iteration, multiple permutation robot simulation
Imitation robot control stack models
The complete family of 8 granted US patents in simulation-driven robotics, plus international filings, assigned to X Development, Google, and Google DeepMind; 11 granted US patents between the two founders in all. Patents are inventor credentials assigned to prior employers.
At Cereal
At Cereal, Matt owns the world loop: the always-on simulation that mirrors the building, verifies every plan before a body moves, and turns recorded operation back into training and verification data.