TechWhoKnows
Log inSign up
Latest
AI
Startups
Venture
Tech
Security
Hardware
Biotech & Climate
Markets & Policy
Companies
TopicsNewsletterAbout
TechWhoKnows

Independent tech news from Singapore, with the sources linked.

Sections

  • AI
  • Startups
  • Venture
  • Tech
  • Security
  • Hardware
  • Biotech & Climate
  • Markets & Policy

TechWhoKnows

  • Companies
  • About
  • Newsletter
  • All topics
  • Search
  • Account

© 2026 TechWhoKnows. All rights reserved.

  • Privacy
  • Terms
Home/Hardware/Robotics & Defence

Google DeepMind Unveils Gemini Robotics 2.5 Vision-Language-Action Architecture for Dexterous Manipulation

DeepMind has released Gemini Robotics 2.5, unifying multimodal reasoning with high-frequency motor control trajectories for industrial robot manipulation.

Google DeepMind Unveils Gemini Robotics 2.5 Vision-Language-Action Architecture for Dexterous Manipulation

By Marcus Vance, Semiconductors & Hardware

14 Nov 2026, updated 16 Nov6 min read

Google DeepMind has unveiled Gemini Robotics 2.5, an advanced vision-language-action (VLA) foundation architecture engineered to control robotic manipulators across complex industrial and domestic environments. The release bridges high-level semantic reasoning with low-level, high-frequency physical motor control, allowing autonomous robots to interpret unstructured natural language commands and execute dexterous manipulation tasks with high adaptability.

Previous robotic foundation models suffered from a fundamental latency mismatch. While large multimodal models excel at visual scene understanding and conversational reasoning, generating autoregressive text tokens is computationally slow, often operating at just 2 to 5 hertz. Industrial robot manipulators, however, require real-time joint velocity and torque updates at 100 to 500 hertz to maintain physical stability, avoid collisions, and manipulate delicate or shifting objects.

Decoupled hierarchical policy architecture

Gemini Robotics 2.5 solves this timing disparity through a hierarchical, two-tier control architecture:

  1. A high-level semantic planner, powered by a fine-tuned Gemini 2.5 multimodal backbone, runs asynchronously at 5 hertz. It analyses continuous video feeds from wrist-mounted and overhead stereo cameras, identifies target objects, formulates task sub-goals, and reasons about physical constraints such as friction, weight, and fragility.

  2. A low-level motor diffusion policy runs synchronously at 200 hertz directly on local edge compute hardware. It translates the high-level semantic sub-goals into smooth, seven-degree-of-freedom joint trajectories and tactile gripper torque outputs.

By decoupling semantic scene decomposition from immediate motor execution, the robot maintains continuous physical responsiveness. If an object slips or an obstacle enters its workspace, the low-level policy adjusts motor trajectories instantaneously without waiting for the high-level multimodal planner to complete a full inference cycle.

Superior zero-shot manipulation of deformable materials

In comprehensive laboratory and factory benchmark evaluations, Gemini Robotics 2.5 demonstrated substantial improvements over predecessor systems like RT-2. The model achieved an 88% overall task success rate across zero-shot manipulation trials, handling tasks it had never encountered during training.

The system showed particular dexterity when manipulating deformable, compliant objects such as cables, cloth, rubber seals, and biological food products. Handling flexible materials has historically proven notoriously difficult for industrial robots because deformable items change shape unpredictably when touched, causing conventional rigid-body kinematic algorithms to fail.

Trained on a diverse mixture of human teleoperation demonstrations, autonomous self-supervised play, and millions of synthetic simulation episodes generated inside Nvidia Isaac Sim and Google MuJoCo physics engines, Gemini Robotics 2.5 demonstrated closed-loop tactile adaptation. When threading a flexible wire through a tight electrical harness conduit, the robot continuously adjusted insertion angles based on real-time force sensor readings.

Edge deployment and industrial commercialisation

To facilitate adoption in commercial manufacturing, automotive assembly, and logistics warehousing, DeepMind designed Gemini Robotics 2.5 to execute inference locally on industrial-grade edge accelerators, including the Nvidia Jetson Thor platform. This local execution capability ensures that factory floor operations continue uninterrupted even during internet connectivity outages.

DeepMind announced partnerships with leading commercial robotics manufacturers, including KUKA, Boston Dynamics, and Universal Robots, to integrate Gemini Robotics 2.5 APIs into standard industrial robot controller interfaces. Factory deployment trials are scheduled to begin in automotive manufacturing and electronics assembly plants across North America and Europe during the first half of 2027.

Frequently Asked Questions

What is a Vision-Language-Action (VLA) model?

A VLA model takes visual imagery and natural language instructions as inputs and directly outputs physical motor actions, such as robot joint angles and gripper torques, allowing robots to perform physical tasks.

Why do robots need high control frequencies like 200 hertz?

High-frequency control is essential for physical stability. If a robot only updates its motor commands 5 times a second, it cannot react quickly enough if an object slips or a collision occurs. Updating commands 200 times a second enables smooth, responsive movement.

Does Gemini Robotics 2.5 require constant cloud connection?

No. The entire model, including its high-level multimodal planner and low-level motor policy, is optimised to run locally on industrial edge computing modules installed directly inside the robot chassis.

Sources

  1. 1DeepMind Robotics Researchdeepmind.google
  2. 2IEEE Spectrum Roboticsspectrum.ieee.org

More on

  • Robotics
  • DeepMind
  • Hardware
  • Automation
  • AI

Further reading

Robotics & Defence

Intel Pushes Core Ultra Series 3 Into Humanoid and Commercial Robotics

16 Nov 2026

Big Tech & Platforms

Hardware Paywalls and Benchmark Manipulation: How Big Tech Locks Features Behind Subscriptions

25 Apr 2026

Robotics & Defence

Why NVIDIA Cosmos and Cadence Just Solved Robotics’ Biggest Problem

18 Apr 2026

Robotics & Defence

Icarus: Inside the $100,000 Solar Stratospheric Drone Startup Challenging Satellites

27 Mar 2026

Follow TechWhoKnows for real-time coverage

Stay updated with the latest developments in technology.

XBlueskyThreadsInstagramLinkedInTikTokYouTube

The TWK Newsletter

Pick your topics and we'll email you our stories on them, with links to the sources.

In this story

Related stories

Intel Pushes Core Ultra Series 3 Into Humanoid and Commercial Robotics

16 Nov 2026

Hardware Paywalls and Benchmark Manipulation: How Big Tech Locks Features Behind Subscriptions

25 Apr 2026

Why NVIDIA Cosmos and Cadence Just Solved Robotics’ Biggest Problem

18 Apr 2026