In late June 2025, Google DeepMind introduced a new AI model named Gemini Robotics On-Device. This model further integrates the multimodal capabilities of Gemini 2.5 directly into terminal devices. Its design focuses on edge or terminal inference, enabling humanoid robots to perform complex tasks by integrating language, vision, and sensory inputs, even in environments without communication or cloud connectivity. This approach marks a significant departure from previous AI systems that relied heavily on cloud computing for processing. Consequently, it enhances the practicality of robots in time-sensitive or network-constrained scenarios such as healthcare and disaster rescue.

Key Highlights

• Humanoid robot technology is evolving from sensing and precise execution towards understanding and autonomous decision-making.
• New offline inference models enable autonomous operation, reducing reliance on remote cloud resources.
• Modularization and edge computing become future trends, enhancing response speed and scene adaptability.
• Development focuses on deployment efficiency and scene integration, rather than solely on linguistic accuracy.
• This trend encourages collaboration between chip designers and application developers, deepening hardware and AI model integration.

Table of Contents

1. Market Update

• Google Unveils Gemini Robotics On-Device Model to Enhance Offline Autonomy of Robots
  • Gemini Robotics On-Device Accelerates Edge Inference as It Is Incorporated into Terminal Devices
    • Figure: Success Rate Comparison between Google Gemini Robotics and Gemini Robotics On-Device
  • Embedded Onboard Models Exert Values of Edge Computing amidst Diversified Development of LLMs for Existing Humanoid Robots
    • Table: Comparison of humanoid robot LLMs in 2Q25

2. TrendForce's View

<Total Pages: 4>