Research Scientist, Robotics Pre-Training and Data Quality, DeepMind

Overview

As a Research Scientist, you will advance robotics foundation models by shaping data quality, acquisition, and labeling strategies, and by optimizing data mixtures. You will oversee training runs, refine recipes, and hill-climb model performance against benchmarks. You’ll build tools to accelerate research, verify data quality through robot policy training, and contribute to initiatives spanning reinforcement learning, vision-language-action modeling, and world-action models. You’ll deploy ideas at scale and collaborate across cross-functional teams to turn research into robust real-world robotics solutions.

Responsibilities

• Drive data quality, acquisition, and labeling strategies for robotics foundation models
• Oversee data mixture optimization for multimodal models used in robotics
• Coordinate and manage training runs; improve training recipes
• Systematically hill-climb model performance against benchmarks
• Build and maintain software tools and processes for rapid iteration on research ideas
• Verify data quality through robot policy training and imitation learning techniques
• Contribute to broader research agenda (RL, vision-language-action modeling, world-action models, simulation)
• Set up large-scale tests and deploy promising ideas quickly and broadly
• Publish findings and share research with the wider community
• Work with simulators and real-world robotic platforms to bridge lab and field applications

Key requirements

• PhD in a technical field or equivalent practical experience
• Experience with algorithmic architectures, data sources, and training and inference techniques for generative multimodal models in robotics (VLAs, WAMs)
• Experience optimizing models and systems (performance tuning, experimentation, debugging)

• collaborative mindset
• results-driven
• strong problem-solving能力
• generative multimodal models (robotics focus)
• robot policy training and imitation learning
• data labeling strategies and data mixture optimization

…