About the Role
You will own how our robots learn new behaviors. That means both the models themselves and the pipeline that produces them: a repeatable path from demonstrations to a policy that works reliably in environments we have never seen.
This is a role for someone who wants their research to run on hardware in the field rather than on a benchmark, and who is early enough in a team's life to set its technical direction.
What You'll Do
Train and ship manipulation policies for production robots.
Own the behavior acquisition pipeline end to end: demonstrations in, reliable policy out, with a measurable cost per new behavior.
Make policies work under contact, where the robot applies and reacts to force, and where locomotion and manipulation cannot be treated separately.
Drive architecture decisions on imitation learning, pretrained representations, and how much to unify across tasks, under a hard constraint that policies run on-device.
Work with Evals to define what success means for each behavior, and use the results to drive the next iteration.
Work with Data Collection to shape what demonstrations get captured, and close the loop when data is the bottleneck.
Build the behavior acquisition pipeline as a platform capability, usable by teams outside our own.
Track where the field is going, and make calls about which directions are worth betting on and which are noise.
What We're Looking For
PhD in robot learning, machine learning, or a closely related field, from a group doing leading work in the area. Equivalent industry research experience at a frontier lab is fine.
Research contributions that others build on: publications at leading venues, or widely used open source work.
Hands-on experience running policies on real robots, not only in simulation or on benchmarks.
Depth in imitation learning: behavior cloning, diffusion or flow-based policies, action chunking, and the failure modes that come with learning from demonstrations.
Strong engineering skills. You build the training infrastructure you need rather than waiting for it.
Judgment about data: what to collect, how much, and when more data is not the answer.
Judgment about generalization, including when a method works because of something specific to your setup rather than because it is sound.
Interest in the unglamorous part of the problem, where a policy that works 70% of the time has to get to 95%.
Bonus Points
Publications at leading venues in robot learning, or well-used open source work.
Experience with contact-rich or whole-body manipulation.
Experience with vision-language-action models or large pretrained policies, including distillation and quantization for on-device inference.
Experience with RL fine-tuning on top of imitation-learned policies.
Internships or research collaborations with industry robotics teams.
Summary
Own how robots learn new behaviors from demonstrations to reliable on-device manipulation policies under contact and whole-body constraints. Requires PhD in robot learning or machine learning or equivalent frontier-lab experience, hands-on real-robot policy deployment, and depth in imitation learning.
Responsibilities
Train and ship manipulation policies for production robots; Own behavior acquisition pipeline end to end; Drive imitation learning architecture decisions for on-device policies; Define success with Evals; Shape demonstration data collection
Qualifications
PhD in robot learning, machine learning, or closely related field; publications at leading venues or widely used open source work; depth in imitation learning; strong engineering skills; judgment about data and generalization
Education requirements
PhD in robot learning, machine learning, or a closely related field; Equivalent industry research experience at a frontier lab is fine
Experience requirements
Hands-on experience running policies on real robots, not only in simulation or on benchmarks; Research contributions that others build on