Research Scientist Intern, Robotics Foundation Models, Agentic RL

About the role#

Meta Reality Labs Research develops world action models and vision-language-action models for robotic manipulation. We are looking for a Research Scientist Intern to focus on scalable offline reinforcement learning for post-training these models. You will conduct research to advance foundation models for robotics, develop new algorithms, and evaluate model capabilities. Internships range from 12 to 24 weeks with various start dates.

What you'll do#

• Develop agentic reinforcement learning algorithms for post-training using large, heterogeneous robot datasets.
• Investigate reward learning, value estimation, and policy optimization for generative action models, including the use of learned world models.
• Solve challenges related to learning from mixed-quality experience, such as distribution shift, sparse rewards, and long-horizon credit assignment.
• Build training and data pipelines while studying how policy performance scales with dataset size, model capacity, and compute.
• Design experiments and benchmarks to test policy improvement and generalization across environments.
• Collaborate with researchers and engineers to evaluate policies in simulation and on physical robots.
• Present research findings and publish results at machine learning and robotics venues.

What you'll need#

• Currently pursuing a Ph.D. in Computer Science, Machine Learning, Artificial Intelligence, Robotics, or a related STEM field.
• Research experience in reinforcement learning, specifically in policy optimization, value learning, or offline/off-policy learning.
• Experience implementing and training deep learning models using Python and PyTorch or a similar framework.
• Ability to design controlled experiments, analyze results, and debug machine learning algorithms.
• Experience working in cross-group and cross-culture environments.
• Intent to return to a degree program after the internship.
• Experience with offline reinforcement learning methods for continuous control or learning from suboptimal experience.
• Experience training or post-training VLAs, WAMs, or other foundation models for embodied decision-making.
• Knowledge of diffusion or flow-based policies, action-sequence prediction, model-based RL, or action-conditioned world models.
• Hands-on experience with distributed training across multiple GPUs or nodes.
• Experience evaluating policies on physical robots or simulation benchmarks.
• A record of results shown through grants, fellowships, patents, or accepted papers in leading conferences or journals.

Location & details#

• Location: Menlo Park, California.
• Modality: On-site.
• Employment: Full-time internship.
• Compensation: $7,650 to $12,667 per month.
• Requirements: Must obtain and maintain work authorization in the United States.