Загрузка видео...
Не удалось загрузить видео
Google DeepMind just introduced Gemini Robotics 2. It's a single VLA that unlocks physical dexterity across different end effectors, hands or grippers, from one model checkpoint. Apptronik's Apollo 2 humanoid, with the 22-DoF SharpaWave hand, ties knots and seals a ziplock bag
133,782 просмотров • 3 месяцев назад •via X (Twitter)
Комментарии: 41

For the first time, a Gemini Robotics model controls an entire humanoid, feet to fingertips, not just the upper body.

The slow thinking brain: Gemini Robotics ER 2 The high-level reasoning model observes the room, plans multi-step tasks lasting minutes (hundreds of decisions), coordinates with the VLA to act, and self-corrects when a step fails. New: multi-robot collaboration

Results: Hits medium-to-high success on whole-body and gripper tasks. But multi-finger dexterity is still hard. On Apollo + Sharpa hands: unscrew bulb 92%, yet tie trash bag 44%, ziplock 40%, dustpan 32%. Dexterous hands remain the challenging frontier.

Availability: Gemini Robotics ER 2 (reasoning): on Google AI Studio, plus private preview on the Gemini Enterprise Agent Platform. The GR2 VLA and lighter On-Device VLA: open to early-access partners.

My conversation with Google DeepMind's Kanishka Rao about the latest model, alternatives to VLAs, the real path to physical AGI, and more:

A good physical Turing test would be: 1. Make a fly 2. Tie it to line 3. Walk to a boat 4. Row a boat 5. Catch a fish 6. Gut a fish 7. Cook a fish 8. Serve a fish And then clean up .

fair

Dope but what happens when it gets bitch slapped?

Computers were useful because they did things quicker, this is like watching paint dry

I've been following Apptronik's Apollo for years - this bot's hardware, except for recent improvements to the hands - hardly seems to have evolved at all. I don't even think Figure AI existed when I first saw Apollo, but progress at Figure appears to have overtaken them.

🤔 meticulously slow...

Ziplock bags were designed for humans, not robots. Let's design the robots to do robot things. I'll seal my own plastic bag thank you.

Checkpoints from the database point of view ??

very slow, why?

Is there something erotic about those gentle movements, or am I just weird? 😂

Yeah, at 100x the amount of time it takes you to do it yourself. Not helpful at all.

Today, George Jetson turns Four. Just saying.

cool trivia!

SharpaWave sounds creepy to me as the name for a hand

Great cut, really highlights the difficulty there. Who could have guessed that this would be possible even 2 years ago?

tying knots and sealing ziplock bags is the true Turing test. imagine a robot trying to get child-proof packaging open. that's the real challenge.

Single VLA across grippers/hands/end-effectors is the actual unlock — one checkpoint, many bodies. Apollo 2 tying ziplock bags is the kind of demo that says the data flywheel started turning. Hard part is teaching 'failure recovery' now.

We're witnessing the operating system for humanoid robots being built in real time. 🤖🚀 What's the first real-world task you'd trust one with?

@elonmusk where's your toy?

A single checkpoint working across different hands and grippers feels like a much bigger milestone than the individual demos. How well does it generalize to hardware it wasn't trained on?

∆

The future of robotics lies in adaptable, general-purpose intelligence. It's exciting to see foundation models enabling more versatile robot capabilities. 🚀🤖

apollo 2 + sharpaWave tying a ziplock is the part nobody is pricing in. dexterity demos used to be scripted; one checkpoint doing multi-embodiment ties the progress back to the underlying VLA, not to the hardware. that's the message worth spreading.

Curious how much of the remaining challenge DeepMind sees as model capability vs. deployment. A stronger VLA doesn’t necessarily make heterogeneous fleet integration any easier.

This stuff is going way to fast... the robotics/ai industry need super regulation and monitoring b4 something bad happens...

Very impressive. A single VLA working across different robot hands and grippers is a big step toward more capable and adaptable robots.

True intelligence reveals itself not in abstract reasoning alone, but in the quiet precision with which thought becomes reliable action amid the messy physics of the real world. Gemini Robotics 2 advances that marriage of mind and body, yet the lingering struggle with simple knots and bags reminds us that the final miles toward embodied mastery will test us most deeply

Curious how it handles latency in real-time control loops.

That's 30 gigs of data for that knot

I just found a new way to tie off my garbage bags! Truly the pinnacle of our civilization!

😮

Can it speed up a little? It was like watching paint dry

The right approach for humanoid robots is to first research the hand nerves.

impressive dexterity, but real-world robustness is the hard part. tying knots in a lab is one thing, handling variable friction and lighting on a factory floor is another. curious how it generalizes beyond controlled demos.

Why can't u just push a fast forward button on these things

Super interesting. From the video they have on youtube I have a hunch its a hybrid VLA. Traditional VLA for long term reasoning and Jepa like for next world state prediction.


