Robotics Intern, Post-Training and Continual Learning

The Team
Our team in the Robotics division is developing pretrained generalist policies that can support a broad range of tasks, environments, and robotic systems. A central scientific challenge is determining how these broadly pretrained policies can be efficiently adapted to new target tasks while achieving the reliability, precision, and robustness required for real-world use.
Pretraining can provide a policy with broad behavioral capabilities, but those capabilities may not immediately translate into dependable performance in a specific deployment setting. Effective post-training methods are therefore essential for converting general capabilities into policies that can adapt quickly, learn from limited additional data or interaction, and perform consistently on demanding downstream tasks.
The Opportunity
Our research interests include reinforcement learning, imitation learning, human-in-the-loop learning, simulation, world models, policy distillation, and large-scale robot learning. We aim to advance the scientific foundations of policy adaptation while developing methods that may ultimately be evaluated or deployed within internal research projects involving real-world industrial tasks.
Potential research directions include, but are not limited to:

• Offline-to-online reinforcement learning

• DAgger, imitation learning, and human-in-the-loop policy improvement

• Sim-to-real policy distillation and adaptation

• Policy improvement, planning, or data generation using world models

• Data-efficient adaptation to new tasks and environments