Google DeepMind unveils Gemini Robotics 2 for whole-body humanoid control and multi-robot teamwork
- Gemini Robotics 2 is DeepMind's new vision-language-action (VLA) model that controls full humanoids from feet to fingertips as well as bi-arm robots, adapting to new robot bodies in a few hours of data.
- The suite also includes Gemini Robotics ER 2, an embodied-reasoning vision-language model that plans multi-step tasks lasting several minutes and lets robots coordinate as a team.
- Gemini Robotics On-Device 2 runs locally on robotic hardware and can adapt to a completely new robot embodiment with just a few hours of data.
- Published benchmarks show wide performance gaps: pick-from-floor tasks hit only 45.7% success while unscrewing a light bulb reaches 92%, but screwing one in drops to 36%, and multi-finger dexterity is flagged as still hard.
- The model adds safety behavior that detects nearby humans and can trigger a safety stop, though commenters note this implies close human-robot interaction still needs a much higher level of control.
Hacker News 의견들
Anyone asking for an honest assessment here: from my prototyping experience, the progress is real but very far from what an average customer would find reliably useful for daily chores. Rodney Brooks' point still stands, dexterity is a hardware problem too, grippers just aren't hands like the paper's own "multi-finger dexterous manipulation remains challenging" line admits.
It's still at GPT-1 level honestly, but the GPT-2 moment feels imminent.
I wrote my bachelor's thesis on VLAs (basically ChatGPT hooked to a robot arm). It's not there for anything practical yet, benchmarks like Libero are saturated so everyone scores 95%, and companies lie a lot about what's teleoperated versus autonomous in their demo videos.
There's a ton of monkey business in robotics demos, sometimes outright fraud claiming autonomy when it's teleoperated. Humanoids are far from useful in real scenarios, and fun fact, most QUGVs can't even reverse on stairs.
Fingers crossed a Chinese lab drops open weights so the rest of us can actually use this stuff.
Qwen already released RobotSuite for this, and Nvidia has the Cosmos series too.
If the robot has to stop whenever a human gets close, that means handling or close interaction with people needs a whole other level of control, right?
Yep, that's a genuinely harder safety problem to solve.
Also notice "better" detection still means it fails regularly, that word is doing a lot of lifting.
A 36% success rate on screwing in a light bulb kind of supports Yann LeCun's argument that VLA/VLM won't be what powers tomorrow's robots.
These robots look slow and clunky now, but LLMs looked dumb early on too. If robotics improves as fast as LLMs did, this gets huge in a few years.
For chores like cleaning, speed barely matters as long as it's done by the time you're home from work.
We don't get the same feedback loops in hardware though. Muscles took half a billion years of evolution to get this good, token generation is maybe a few hundred thousand years of progress by comparison.
GPT-2 to GPT-3 was 15 months and the jump was massive. Gemini Robotics 1 to 2 was also 15 months apart but barely changed, and fast fluid motion is a totally different difficulty tier from slow careful walking.
Isn't 'if progress is as fast as LLMs' just the new version of 'self-driving is a year away'?