Principal Scientist, Physical AI

Your New Role and Team

Sanctuary, a world leader in building AI-based control systems for intelligent robots, is seeking a Principal Scientist to support our next generation of robot learning capability. This role sits at the intersection of deep technical understanding and rigorous validation: you will push the boundary of what robots can learn to do, and you will run the experiments that show which approaches have real merit and are worth scaling into a production capability. Deep expertise in at least one or more areas of robot learning — RL and imitation learning, simulation and sim2real, data collection, or world and foundation models — and a working knowledge or awareness of the rest. 

This role begins with reviewing everything we are doing and introducing new approaches: you will scope and run focused, high-leverage projects, and bring in support as your work proves out. Successful projects will drive larger projects and more resources.

Our Success Criteria

• Identify, scope, and lead research initiatives, working with contributors from across the company

• Design and run experiments that validate early hypotheses, surfacing what works, what does not, and what deserves further investment

• Demonstrate the merit of new approaches through hands-on experimentation, building the evidence base that demonstrates readiness to scale toward production

• Ability to create, develop, and enhance cutting-edge robot learning algorithms — RL, imitation learning, world models, foundation models, or approaches we have not tried — and evaluate their performance in practical robotic applications

• Devise training and data collection pipelines to expedite implementation on physical robots

• Discover strategies for enhancing current learning processes, considering key performance metrics like sample efficiency, speed, computational resources, and scalability

• Collaborate within a diverse team to devise, implement, and harden algorithms for production use, and investigate the root causes in existing implementations

• Translate Machine Learning research and trained models into real-world robotic products, working closely with engineering to validate that promising methods hold up as they move toward our cross-platform software framework

• Stay current with the latest developments in robot learning — including RL/IL, world models, and foundation models — and their application in robotics

Your Experience

Qualifications

• Ph.D. in Machine Learning, Computer Science, Applied Mathematics, or equivalent practical background in robot learning

• 5+ years of hands-on experience implementing and deploying robotic manipulation tasks, both in simulation and on physical robots

• Experience transitioning Machine Learning research and trained models into real-world production

• Active involvement integrating Machine Learning models into a robotics platform

• Strong technical judgment and the ability to form and defend an independent point of view

• Independently able to evaluate a landscape, propose an approach, complete a proof-of-concept and document and articulate the results to a cross-functional team

• Comfortable challenging assumptions and driving alignment across technical stakeholders

Deep expertise in at least one of

• Applying various Reinforcement Learning and/or Imitation Learning methods, with focus on robotics in the real world

• Developing and optimizing large-batch parallel simulations, and proven expertise in sim-to-real transfer

• Designing and scaling data collection workflows and datasets for robot learning on physical systems

• Building or fine-tuning world models or foundation models for robotics

Nice to Have

• Proven expertise in continual learning, employing adaptive model training to improve long-term performance and accuracy

• Research contributions at venues such as ICRA, IROS, CoRL, or NeurIPS, or open-source work

Skills

• Development with Python 3.8 or later

• Working knowledge of PyTorch and/or TensorFlow

• Familiarity with ROS2

• Strong understanding of modern robot learning methods and their application

Traits

• Above all else, a consistently positive attitude and a willingness to do whatever it takes to create robust solutions to complex problems

• Pragmatic and outcome-oriented, with a bias toward turning advanced research into real-world products

• Entrepreneurial: comfortable scoping and running a project with a cross-functional team 

• Strong leadership skills in organizing R&D work for projects, with the ability to lead a cross functional team 

• Curious, adaptable, and willing to go deep into unfamiliar technical areas when needed

• Patience, persistence, and attention to detail when resolving performance issues

• Ability to multitask and prioritize in a fast-paced environment