AI foundation models give robots the ability to see, reason, and act in the physical world

Overview

For decades, industrial robots have been powerful but rigid—they follow pre-programmed instructions and break when the world deviates from what engineers anticipated. Google DeepMind's Gemini Robotics-ER 1.6, released on April 14, 2026, represents the sharpest version yet of a new approach: giving robots the same kind of flexible reasoning that powers chatbots like Gemini, but aimed at understanding and acting in physical space. The model can now read pressure gauges, thermometers, and digital readouts with 98% accuracy when paired with its agentic vision system—a capability that emerged directly from collaboration with Boston Dynamics, whose Spot robot patrols industrial facilities.

This release is one move in a broader race to build what the industry calls "physical AI"—foundation models that let any robot understand novel environments without being explicitly programmed for them. Google DeepMind, NVIDIA, and a growing roster of Chinese manufacturers are all converging on the same bet: that the architectures behind large language models can be adapted to give robots general-purpose intelligence. The stakes are enormous. The humanoid robot market is projected to grow from $2 billion in 2024 to over $13 billion by 2029, and the companies that control the underlying AI models will shape which robots can do what.

Why it matters

The AI models that power chatbots are now learning to control physical robots, setting the stage for autonomous machines in factories, warehouses, and homes.

Play on this story Voices Debate Predict

Key Indicators

98%

Gauge-reading accuracy

Gemini Robotics-ER 1.6 with agentic vision achieves 98% accuracy reading industrial instruments, up from 23% in the previous version

$13B+

Projected humanoid robot market by 2029

The market is expected to grow from $2 billion in 2024 at a 45% compound annual growth rate

13,317

Humanoid robots shipped globally in 2025

Chinese manufacturers claimed 87% of unit volume, led by Unitree Robotics with over 5,500 units

30,000/yr

Boston Dynamics Atlas production target

Factory-scale production target for the Atlas humanoid robot, powered by Gemini Robotics foundation models

Interactive

Exploring all sides of a story is often best achieved with Play.

Ever wondered what historical figures would say about today's headlines?

People Involved

Demis Hassabis

Leading Google's push into physical AI

Robert Playter

Overseeing Gemini integration into Spot and Atlas platforms

Organizations Involved

Google DeepMind

Technology platform; AI assistant operator

Releasing foundation models for robotics at accelerating pace

Google's primary artificial intelligence research division, now building the Gemini Robotics model family to extend large language model capabilities into physical robot control.

Boston Dynamics

Robotics Company

Integrating DeepMind AI into Spot and Atlas platforms

A Hyundai-owned robotics company known for its Spot quadruped and Atlas humanoid, now partnering with Google DeepMind to add AI reasoning capabilities to its hardware platforms.

Nvidia Corporation

Semiconductor company

Building competing open-source robotics foundation models

The dominant supplier of AI training hardware, now competing directly with Google DeepMind by releasing open-source robotics foundation models through its Isaac GR00T platform.

Timeline

Google DeepMind releases Gemini Robotics-ER 1.6
Release

The new model significantly improves spatial and physical reasoning—pointing, counting, and success detection—and introduces instrument reading, enabling robots to autonomously read pressure gauges, thermometers, and digital displays with up to 98% accuracy using agentic vision. Available immediately via the Gemini API and Google AI Studio.
April 14th, 2026
Hassabis predicts robotics "ChatGPT moment" at Davos
Statement

Google DeepMind's CEO told the World Economic Forum that the field would witness a breakthrough in physical intelligence within 18 months, comparable to the public impact of ChatGPT.
January 20th, 2026
Boston Dynamics and Google DeepMind announce partnership at CES 2026
Partnership

At Hyundai's CES press conference, the two companies revealed plans to integrate Gemini Robotics foundation models into both the Spot quadruped and the next-generation Atlas humanoid, with a factory production target of 30,000 Atlas units per year.
January 5th, 2026
Gemini Robotics 1.5 released with expanded capabilities
Release

The updated model improved spatial understanding, natural language interaction, and multi-step task planning, and added the ability to call external tools like Google Search during operation.
September 1st, 2025
Gemini Robotics On-Device released for low-latency local control
Release

A lightweight version of Gemini Robotics optimized to run directly on robot hardware, enabling real-time dexterous tasks like unzipping bags and folding clothes without cloud connectivity.
June 1st, 2025
Google DeepMind introduces Gemini Robotics and Gemini Robotics-ER
Release

DeepMind debuted two models built on Gemini 2.0: Gemini Robotics, a full VLA model that directly controls robots, and Gemini Robotics-ER, a reasoning-focused model for environmental understanding and task planning.
March 12th, 2025
Boston Dynamics retires hydraulic Atlas, unveils electric version
Hardware

Boston Dynamics replaced its famous hydraulic humanoid with an all-electric design built for commercial manufacturing, signaling a shift from research spectacle to factory-floor deployment.
April 17th, 2024
NVIDIA announces Project GR00T for humanoid robots
Competition

NVIDIA entered the robotics foundation model race, announcing a dedicated platform for training humanoid robot AI alongside major hardware and simulation investments.
March 18th, 2024
DeepMind publishes Robot Constitution safety framework
Safety

As part of its AutoRT system, DeepMind introduced a set of safety rules governing AI-controlled robots, inspired by Asimov's laws but designed for real-world deployment constraints.
January 4th, 2024
Google DeepMind releases RT-2, the first vision-language-action model
Research

RT-2 demonstrated that a robot could benefit from training on internet-scale text and image data, translating web knowledge into physical actions. It established the vision-language-action (VLA) paradigm that all subsequent robotics foundation models have followed.
July 28th, 2023

Scenarios

Predict which scenario wins. Contrarian picks score more — points lock in when the scenario resolves.

Predictions leaderboard

Physical AI becomes standard infrastructure in industrial facilities by 2028

AI-powered robots like Spot become routine in factories, refineries, and warehouses, handling inspections, material transport, and hazard detection autonomously. Google DeepMind and NVIDIA compete for the platform layer the way Android and iOS competed for smartphones. Instrument reading, spatial reasoning, and multi-step task planning reach human-level reliability. This scenario is accelerated by the Boston Dynamics partnership and Hyundai's manufacturing capacity.

Discussed by: Deloitte's 2026 Tech Trends report, NVIDIA investor briefings, Boston Dynamics commercial roadmap

Consensus —

China dominates physical AI through hardware volume while the West controls models

Chinese manufacturers like Unitree and Agibot continue to ship the vast majority of humanoid robot hardware at price points Western companies cannot match ($5,900 to $13,500 per unit). But Western companies—Google DeepMind, NVIDIA, OpenAI—control the foundation models that make these robots useful. A split emerges: Chinese hardware running American AI, creating complex supply chain dependencies and geopolitical tension similar to the current semiconductor dynamic.

Discussed by: Rest of World reporting on Chinese robotics dominance, Goldman Sachs humanoid robot analysis

Consensus —

Foundation model robotics hits a capability wall, deployment stalls

Despite impressive demos, foundation model robots struggle with edge cases in unstructured real-world environments—the "last 2%" of reliability that separates demonstrations from dependable industrial tools. Safety incidents during early deployments trigger regulatory scrutiny. Enterprise customers pull back, and the 45% growth projections prove premature. The technology eventually works but on a slower timeline than current hype suggests.

Discussed by: Robotics researchers at MIT and Carnegie Mellon, skeptical analysis from IEEE Spectrum

Consensus —

Humanoid robots enter homes by 2028, reshaping domestic labor

If foundation models advance fast enough and hardware costs continue falling, general-purpose humanoid robots priced under $20,000 begin appearing in households for elder care, cleaning, and domestic tasks. Tesla's Optimus, initially deployed in Tesla factories, becomes the first mass-market home robot. This scenario depends on solving both the AI reliability problem and consumer trust—neither is guaranteed.

Discussed by: Tesla Optimus roadmap, Unitree consumer marketing, Elon Musk public statements

Consensus —

Historical Context

Unimate and the birth of industrial robotics (1961)

1961

What Happened

George Devol's Unimate, a 4,000-pound robotic arm, was installed at General Motors' die-casting plant in Ewing, New Jersey—the first industrial robot deployed in a factory. It performed a single task: lifting and stacking hot metal parts that were dangerous for human workers. The machine cost the equivalent of $500,000 in today's dollars and could only follow pre-programmed magnetic drum instructions.

Outcome

Short Term

GM expanded robotic welding across its plants within five years. By 1969, the Stanford Arm introduced computer-controlled articulation, making robots more flexible.

Long Term

Industrial robots became foundational to manufacturing. By 2024, over 4 million industrial robots were operating worldwide, but nearly all still required explicit programming for each specific task.

Why It's Relevant Today

Gemini Robotics-ER 1.6 represents the next transition: from robots that execute pre-programmed instructions to robots that reason about novel situations. Just as Unimate proved machines could handle dangerous factory work, foundation model robots are proving machines can handle unpredictable factory environments.

ImageNet and the deep learning revolution (2012)

October 2012

What Happened

Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton entered a deep neural network called AlexNet into the ImageNet Large Scale Visual Recognition Challenge and crushed the competition, reducing image classification error from 26% to 16%. The result stunned the computer vision community and triggered a rush to apply deep learning to every perceptual task.

Outcome

Short Term

Within two years, every major tech company had launched or acquired a deep learning research lab. GPU maker NVIDIA's stock began a decade-long ascent.

Long Term

The ImageNet moment proved that scale plus data plus compute could solve problems that hand-engineered approaches had struggled with for decades. It directly led to the transformer architectures behind today's language models and, now, robotics foundation models.

Why It's Relevant Today

Demis Hassabis has explicitly compared the current state of robotics AI to the period just before deep learning's breakout. The pattern is similar: a general-purpose architecture (transformers/VLAs) is being applied to a domain (physical robotics) that previously required domain-specific engineering. Gemini Robotics-ER 1.6's jump from 23% to 98% instrument-reading accuracy echoes the kind of sudden capability gains that AlexNet delivered for image recognition.

Kiva Systems and warehouse automation (2012-2015)

March 2012 - 2015

What Happened

Amazon acquired Kiva Systems for $775 million in March 2012, gaining a fleet of small orange robots that could autonomously navigate warehouse floors to bring shelving units to human pickers. Amazon rebranded the division as Amazon Robotics and stopped selling the technology to competitors, forcing rivals like Walmart and FedEx to develop or acquire their own solutions.

Outcome

Short Term

Amazon deployed over 200,000 Kiva robots across its fulfillment centers within three years, significantly reducing order processing time.

Long Term

Amazon's robot acquisition triggered a warehouse automation arms race. By 2026, the warehouse robotics market exceeded $9 billion annually, and companies like Agility Robotics were deploying humanoid robots in warehouse settings.

Why It's Relevant Today

The Kiva acquisition showed how a single company's AI robotics investment could reshape an entire industry's competitive dynamics. The Google DeepMind–Boston Dynamics partnership carries similar potential: if Gemini-powered robots prove significantly more capable than competitors, the companies that control the foundation models gain outsized influence over industrial automation.