EXPLORINEER
en

Archive — US — techcrunch

Google rolls out new Gemini model that can run on robots locally

Google has launched a new local language model, Gemini RoboticsOn-Device, enabling robots to perform tasks without an internet connection. Building on March's Gemini Robotics model, it offers natural language control and excels in benchmarks. It has proven effective on various robots, including the bi-arm Franka FR3, to automate tasks like assembly and clothing folding.

Key points

  • Google DeepMind's Gemini RoboticsOn-Device can run tasks locally on robots.
  • The model builds upon the previous Gemini Robotics model released in March.
  • It allows developers to fine-tune robot movements using natural language prompts.
  • The model performs comparably to cloud-based versions and outperforms other on-device models.
  • In demonstrations, it successfully controlled robots to unzip bags and fold clothes.
  • The model adapts for different robot types, showcasing versatility in tasks.
  • The Gemini Robotics SDK facilitates developer training through task demonstrations.
  • Competitors like Nvidia and Hugging Face are entering the robotics AI space.

Why it matters

  • The rise of on-device AI models aligns with increasing concerns over data privacy and latency.
  • Robotics is becoming a significant field within AI, with numerous startups and tech giants investing in foundational models.
  • The 2025 IT trends indicate a shift toward decentralized computing and autonomous systems in various industries.

Google's Gemini RoboticsOn-Device represents a significant advancement in local AI capabilities for robotics, reflecting broader trends toward decentralized systems and prompting competition among tech companies in the robotics field.