Research Scientist, Reinforcement Learning & Robotics

The work

A companion needs to understand a situation and take useful actions as it changes. This role develops learning methods for policies in games and simulated environments, drawing on reinforcement learning and robotics to study control, planning, and adaptation. The goal is to build policies that can learn from demonstrations, improve through experience, and work alongside people.

What you’ll work on

• Develop policies that combine imitation learning with offline and online reinforcement learning, using human demonstrations and experience collected through play.
• Build simulation environments, training curricula, and evaluation tasks for control, exploration, planning, and cooperation with human players.
• Research world models and hierarchical policies that connect observations and long-term goals to actions under partial observability.
• Study how policies generalize across environments, tasks, and control interfaces, including adaptation methods informed by sim-to-real research.
• Design rewards, collect training experience, and run controlled experiments that separate improvements in task performance from changes in data or evaluation conditions.
• Implement efficient training and inference systems, and connect learned policies with the companion’s perception, memory, and conversational behavior.

What you’ll bring

• A PhD in robotics, machine learning, computer science, or a related field from a leading research university, or an equivalent record of original research.
• A strong publication record at major conferences such as NeurIPS, ICLR, ICML, CoRL, RSS, ICRA, or IROS.
• Deep understanding of reinforcement learning and sequential decision-making, with experience developing and evaluating learned policies.
• Strong Python and PyTorch skills, including experience implementing training algorithms, debugging optimization, and building reproducible experiments.
• Experience working with simulation, embodied learning, robotic control, or interactive environments, and the judgment to design meaningful evaluation protocols.

Helpful experience

• Research on imitation learning, offline RL, model-based control, world models, or learning from human feedback.
• Experience with sim-to-real transfer, domain randomization, system identification, or adaptation across tasks and embodiments.
• Experience with multi-agent learning, human-AI cooperation, or distributed training in games and simulation.