Research Scientist Intern, Robotics Foundation Models Mid-Training (PhD)

About the job

Meta Reality Labs Research is a team of researchers and engineers pushing the frontier of AI and robotics. Our team is building the next generation of world action models (WAMs) and vision-language-action models (VLAs) for robotic manipulation. We are seeking a Research Scientist Intern to develop scalable methods for mid-training world action models on large-scale egocentric video and robot data. Interns in this role will conduct fundamental and applied research to advance foundation models for robotics, with a focus on mid-training. You will collaborate with research scientists and engineers to develop new algorithms, build and evaluate model capabilities, and contribute to the next generation of embodied AI. Our internships are twelve (12) to twenty-four (24) weeks long and we have various start dates throughout the year.

Responsibilities

Develop scalable mid-training methods for WAMs using large-scale egocentric video and heterogeneous robot data. Investigate representation and objective design for learning from human egocentric video. Address challenges in training on web-scale, weakly-labeled data, including embodiment and viewpoint gaps, data curation and filtering, and domain mixture weighting. Build efficient training and data pipelines, and study how downstream policy performance scales with dataset size, model capacity, and compute. Design experiments and benchmarks to measure what mid-training contributes, isolating its effect on downstream post-training and generalization across tasks and embodiments. Collaborate with research scientists and engineers to evaluate mid-trained models in simulation and on real robots. Communicate research findings and publish results at leading machine learning, computer vision, and robotics venues.