Google DeepMind

Google DeepMind's Robotics Research

Google DeepMind

Home > Companies > Google DeepMind


Overview

Google DeepMind is a pioneer in robotics AI, leading key advancements in VLA models from RT-1 to Gemini Robotics. The current organization was formed in 2023 through the merger of Google Brain and DeepMind.

ItemDetails
HeadquartersLondon, UK / Mountain View, CA
Founded2010 (DeepMind), 2023 (merger)
CEODemis Hassabis
Robotics LeadVincent Vanhoucke

Key Models

Timeline

YearModelKey Contribution
2022RT-1First large-scale Robotics Transformer
2023RT-2First VLA, “Action as Language”
2023RT-XOpen X-Embodiment (33 research labs collaboration)
2025Gemini RoboticsGemini 2.0-based VLA

RT-1 (2022)

  • Trained on 130K episodes
  • Generalized to 700+ tasks using Transformer architecture
  • 97% success rate (on trained tasks)

RT-2 (2023)

  • “First VLA” - Outputs actions as language
  • Based on PaLM-E/PaLI-X (55B)
  • Emergent capabilities: Understanding abstract commands

RT-X (2023)

  • Collaboration of 33 research labs (Open X-Embodiment)
  • 22 robot types, 160K+ episodes
  • Open dataset release

Gemini Robotics (2025)

  • Most powerful VLA based on Gemini 2.0
  • 2x+ performance over SOTA
  • On-Device version released (local execution)

Research Approach

Foundation Model Method

[Large-scale Pre-training] → [Fast Fine-tuning] → [New Task Adaptation]

Core Principles

  1. Scaling: More data, larger models
  2. Multimodal: Integration of Vision + Language + Action
  3. Cross-robot: Training/deployment across diverse robots
  4. Collaboration: Open research like Open X-Embodiment

Partnerships

PartnerDetails
ApptronikApplying Gemini 2.0 to Apollo humanoid
Boston DynamicsTesting Gemini Robotics-ER
Agility RoboticsDigit robot collaboration

References


See Also