Robotics
Why Robots Still Can't Pick Up a Grape Without Crushing It
The most sophisticated humanoid robots on earth can perform backflips but routinely fail at tasks a toddler masters before age two.
The Idea
Touch is the sense robotics has never quite solved. Vision, hearing, even smell — engineers have made remarkable progress translating these into machine-readable signals. But the sense of touch involves something deceptively complicated: not just detecting contact, but reading the texture, compliance, temperature, and shifting pressure of an object in real time, then feeding that information back to adjust grip force within milliseconds. Humans do this without thinking. Robots do not do it at all, not really. The deeper problem is what researchers call the 'embodied cognition gap.' Our hands are not just tools our brains control — they are part of how we think. The fingertip contains roughly 2,500 mechanoreceptors per square centimetre, each tuned to a different quality of contact. When you pick up a grape, you are running a continuous loop: feel, assess, adjust, feel again. The brain and hand are co-processing the problem, not operating in sequence. Most robotic grippers work on a fundamentally different logic: plan the grip in advance, execute it, hope for the best. This works beautifully in structured environments — a factory floor where every object is in the same place, every time. It fails spectacularly in the messy, variable, unpredictable world humans actually inhabit. The frontier of robotics right now is not strength or speed — it is the ancient, unglamorous problem of how to hold something gently.
In the World
In 2022, a team at MIT's Computer Science and Artificial Intelligence Laboratory unveiled a robotic hand equipped with a sensor called GelSight — a small camera embedded beneath a soft gel fingertip that reads deformation patterns to infer texture and force. The hand could identify objects by touch alone, in the dark, with an accuracy rivalling human performance on the same test. It made headlines. It was genuinely impressive. And yet, when the same system was asked to unpack a bag of groceries — the kind of task a ten-year-old does while distracted by a conversation — it struggled. Soft objects shifted. Packaging crinkled unpredictably. Items were stacked, occluded, wedged. The sensor worked; the problem was that touch data alone is not enough. You need touch integrated with vision, with memory of how similar objects behaved before, with an intuition about what is likely to happen next. You need, in short, something closer to experience than to sensing. This is why Amazon's warehouse robots — among the most advanced deployed at scale — still hand off to human workers for the final step of picking irregular items from shelves. The robot is faster, stronger, and tireless. But for the last few centimetres of the job, a human hand remains irreplaceable. That gap, small as it sounds, is where billions in research funding is currently aimed.
Why It Matters
This is not just a robotics problem — it is a window into what intelligence actually is. We tend to assume that the hardest cognitive tasks are things like playing chess or writing poetry, and that physical manipulation is simple by comparison. Decades of robotics research have quietly inverted that assumption. Chess was solved. Folding a fitted sheet has not been. Hans Moravec noticed this in the 1980s and gave it a name — Moravec's Paradox — but it still surprises people, because we judge difficulty by how much conscious effort something takes. Gripping a grape feels effortless, so we assume it is easy. But ease of experience is the wrong measure of computational complexity. Some of the richest intelligence in our bodies runs entirely below awareness. That reframe has practical consequences. It should make you sceptical of predictions about which jobs robots will take first — the answer is consistently 'not the ones involving unstructured physical dexterity.' It should also leave you with a quiet appreciation for what your hands are doing right now, without being asked.
A Question to Ponder
If the hardest problems to automate are the ones that feel effortless to us, what does that suggest about where human value actually lives?
Get a new one of these every morning.
Start learning with Thinkable