Research Scientist

About The Job

Google DeepMind is a pioneering AI lab focused on advancing AI development to solve complex global challenges and accelerate product innovation. Our mission is to combine Gemini foundation models with physical intelligence for general-purpose robots to understand and act reliably in the real world. This involves obtaining high-quality training data and integrating it effectively into the Gemini training process. Our team researches how to best extract physical and action knowledge from robot-related datasets and incorporates these insights into Gemini pre-training through model training and upstreaming. By pairing data ablations with fast closed-loop robot evaluations, you will shape how physical reasoning and action capabilities emerge in next-generation Gemini models and deploy onto real-world robots.

Key Responsibilities

• Develop data quality, filtering, and labeling methodologies.
• Optimize multimodal data representations and mixtures to upstream physical and action knowledge into Gemini pre-training.
• Design and run pre-training and policy training ablations.
• Improve training recipes and systematically hill-climb physical reasoning and action performance against benchmarks.
• Build and maintain research software, evaluation tools, and lightweight post-training workflows for rapid iteration and verification of upstream data quality through offline and closed-loop robot policy evaluations.
• Contribute to the research agenda, including vision-language-action (VLA) modeling, world-action models, reinforcement learning, and simulation.
• Communicate findings clearly in a fast-paced, collaborative team environment.
• Setup large-scale tests and deploy promising ideas quickly and broadly, managing deadlines and deliverables.
• Apply the latest theories to develop new and improved products, processes, or technologies.
• Actively contribute to the wider research community by sharing and publishing findings.

Requirements & Tech Stack

• Experience working with robot policy learning (e.g., imitation learning, RL).
• Experience across simulators and real-world robotic platforms (e.g., dexterous manipulation, multimodal sensing).
• Experience working with simulators and robots.
• Experience with systematic benchmarking.
• Experience with multimodal data sources.
• Experience with data labeling strategies.
• Experience with data mixture optimization for frontier or robotics foundation models.
• Practical experience in "hill-climbing" model performance (e.g., iteratively improving data curation, training recipes, and model benchmarks).
• Proven ability to build and maintain clean research software and tools for rapid iteration on research ideas.
• A passion for bringing foundation model research from the lab to robust, real-world robotic systems.

Qualifications

• PhD in Computer Science, a related technical field, or equivalent practical experience.
• Experience contributing to research communities or efforts, including publishing papers at conferences (e.g., NeurIPS, CoRL, ICML, ICLR).