Google's robots learn to 'think' first

PLUS: Patients control AI and robotics with thought

Jennifer Mossalgue

September 29, 2025

^{Read Online}^|^{Sign Up}^|^Advertise

Good morning, robotics enthusiasts. Google DeepMind just gave robots an internal monologue.

The company’s new Gemini Robotics 1.5 vision-language-action model lets machines explain their reasoning in real time as they work. Consider it Google’s answer to rivals like Figure’s Helix and Nvidia’s Groot.

In today’s robotics rundown:

DeepMind’s robots learn to think aloud
Meta wants to be the ‘Android of robotics’
Unitree’s Bluetooth backdoor nightmare
iRobot founder: humanoid hype is fantasy
Quick hits on other robotics news

LATEST DEVELOPMENTS

GOOGLE DEEPMIND

🧠 DeepMind’s robots learn to ‘think’ aloud

Image source: Google DeepMind

The Rundown: Google DeepMind just released Gemini Robotics 1.5 and Gemini Robotics-ER 1.5 — AI models designed to make robots ‘think’ before they act, translating visual input and language into coordinated movement.

The details:

Google says that the breakthrough lets robots reason through multistep tasks and actually explain what they're doing while they do it.
The system can tackle tasks like sorting recyclables by searching guidelines online, then planning how to physically categorize items based on what it finds.
Unlike traditional robots that react to commands, it generates an internal reasoning process in natural language, breaking complex tasks into steps.
Gemini Robotics-ER 1.5 orchestrates high-level strategy and calls digital tools, while Gemini Robotics 1.5 converts those plans into precise motor commands.

Why it matters: DeepMind’s system hit state-of-the-art performance across 15 robotics benchmarks and works across a wide range of platforms, from dual-arm lab bots to humanoids. The pitch to industry? One adaptable software stack that can power any robot form factor, straight out of the lab and into the field.

🔥 Meta wants to be the ‘Android of robotics’

Image source: Jeff Sainlar, Meta, Wikimedia Commons

The Rundown: Meta is building a software backbone for humanoids. With Project Metabot, it’s pouring billions into AI that hardware makers can license — turning its Llama-powered software into a standard brain for next-gen humanoids.

The details:

CTO Andrew Bosworth told The Verge that Meta is working on the biggest challenge for humanoids: dexterous hand manipulation.
While robots can perform stunts, tasks like grasping a water glass are unsolved because of complex sensor loops and the absence of robust world models.
Meta’s strategy is to leverage its AI Superintelligence Lab to build the kind of simulation, training datasets, and control software needed for fine manipulation.
The team, led by former Cruise CEO Marc Whitten and MIT’s Sangbae Kim, focuses on creating licensable software rather than manufacturing hardware.

Why it matters: Instead of competing with Tesla’s Optimus or Figure on humanoid exteriors, Meta is stacking its strategy with top-tier talent from AV and AR, betting that the real race is in universal manipulation skills. If Meta cracks dexterity (no small feat), it positions itself as the software foundation every robotics manufacturer needs.

UNITREE

🏴‍☠️ Unitree’s Bluetooth backdoor nightmare

Image source: Unitree

The Rundown: Unitree's robot dogs and humanoids are caught in a security nightmare after researchers exposed UniPwn, a wormable Bluetooth exploit that hands attackers root-level control with embarrassing ease.

The details:

The flaw exploits hardcoded AES keys in Unitree's Go2, B2, G1, and H1 robots — attackers simply encrypt the string "unitree" to bypass authentication.
Malicious code disguised as Wi-Fi credentials executes with root privileges, no validation required.
It's wormable: infected robots autonomously scan for and compromise nearby Unitree machines via Bluetooth, creating self-spreading botnets.
Researchers reportedly disclosed the vulnerability in May, but Unitree went silent after July, while UK police are already testing vulnerable Go2 units.

Why it matters: This kind of vulnerability turns advanced robots into potential weapons — spying, sabotaging, or spreading malware autonomously. With police and researchers already field-testing affected units, Unitree’s alleged silence raises serious concerns about security standards in the fast-moving robotics industry.

IROBOT

🧞‍♂️ iRobot founder: humanoid hype is fantasy

Image source: Christopher Michel, Wikimedia Commons

The Rundown: Investors are piling billions into humanoid startups, but Rodney Brooks — a legend who built iRobot and shaped robotics at MIT — has a blunt message: wake up, you're pouring money into a fantasy.

The details:

In his latest essay, Brooks skewers the industry’s big bet on teaching robots by showing them endless human task videos, dubbing it “pure fantasy thinking.”
He ridicules the idea that video-based machine learning can train robots to mimic human dexterity, noting that human hands have 17K touch receptors.
Machine learning's breakthroughs in speech and vision relied on decades of established recording tech — robotics has no such foundation for tactile data.
He warns that scaling up humanoids creates exponential safety risks: doubling a robot's size multiplies its impact energy eightfold.

Why it matters: Brooks predicts that within 15 years, the most successful ‘humanoids’ won't look human at all — they'll use wheels, multiple arms, and specialized sensors rather than walking upright, while current funding rounds will largely evaporate without producing mass-market machines. Of course, Figure and Tesla won’t likely agree.

QUICK HITS

📰 Everything else in robotics today

Major League Baseball will roll out robotic umpire tech in 2026, letting teams challenge two-ball or strike calls a game with instant reviews via Hawk-Eye cameras.

Skild AI says its “omni-bodied robot brain” can universally control any robot — trained on 100K configurations, it adapts to extreme damage like chainsawed-off limbs.

A former Tesla employee is suing Tesla and Fanuc for $51M after a robotic arm allegedly struck and knocked him unconscious while working at the Fremont factory.

University of Virginia scientists developed HydroSpread, a method that lets them fabricate soft robots directly on water by spreading ultra-thin polymer films.

China surged ahead of the U.S. and the world in factory automation, installing nearly 300K new robots last year and bringing its total to over 2M, a new report says.

Beijing’s Horizon Robotics raised about $821M through a Hong Kong top-up placement, offering 639M shares at HK $9.99 each to fund its international expansion.

NYK Line expanded its partnership with Neptune Robotics to deploy robotic hull cleaning across its global fleet for major fuel savings and maritime decarbonization.

Researchers created an AI system that builds on past emotional experiences while learning new ones, making robots more emotionally aware in human interactions.

Hyundai's air taxi startup, Supernal, saw its chief strategy officer, safety officer, and chief of staff depart weeks after pausing its program and losing its CEO and CTO.

COMMUNITY