Can AI take control of a robot?
The AI researchers at Andon Labs, the people who gave Anthropic Claude an office vending machine to run, and hilarity ensued, have published the results of a new AI experiment.
They wanted to see if LLMs were technically capable of functioning as a robot’s brain, that is, connecting their “thinking” (textual decision-making) with real sensors and movement.
The result:
➡️ LLMs are not yet ready to safely and reliably control robots. – even the best models scored below 40% accuracy, humans reached 95%.
➡️ They made funny but also dangerous mistakes (falling down stairs, misinterpreting the environment, or reacting chaotically).
So I wonder how long it will take to achieve something like this, when Elon Musk wanted to present his robots as autonomous units? (and even then, it turned out that they were not completely independent)
How do you see the future of AI integrated into hardware robots?
Replies
I actually believe this transition - AI controlling real-world robots - will happen faster than most anticipate.
Web and desktop AI are relatively easy: publish, iterate, shut down if things go wrong. With hardware it’s different. It’s more expensive, you must ship higher quality, deal with physical safety & logistics.
That means we’re still in a quiet phase. But when the break-through happens, I think we’ll see a flood of market releases - many hardware+AI solutions launching in parallel.
So yes, it may seem “quiet now”, but that’s exactly what precedes the moment when it all scales.
minimalist phone: creating folders
@shahar_shalev In how many years do you think AI will control hardware robots?
TinyCommand
This is a fascinating milestone, but also a necessary reality check.
LLMs are incredible at interpreting context but still struggle with situational awareness something robotics depends on entirely. The gap lies in sensor fusion and real-time adaptability, not reasoning.
I think the future of AI-powered robots won’t come from giving an LLM full control, but from hybrid systems where symbolic reasoning handles structure, and AI models handle perception or language.
We’re probably a few breakthroughs away from true autonomy, but experiments like this are exactly how we’ll get there. Curious if Andon Labs shared how they handled safety overrides during the tests?
minimalist phone: creating folders
@priyanka_gosai1 Maybe AI is not so prepared for real-life conditions and environment. But learns quickly so enough training can make a huge difference. I think that I didn't read anything about safety. Maybe I missed it.
TinyCommand
@busmark_w_nika That’s a great point!
you’re right, training and iteration can close the gap faster than we think. Real-world feedback loops are where models truly evolve.
And yes, I didn’t find much detail about safety either, I hope they’ll share more on how they prevented harm or managed edge cases. That part usually tells us how close we really are to practical deployment.
Present models are not so good.
minimalist phone: creating folders
@joseph84 What do they lack?