A look at how robotic hands are evolving to perform more complex tasks on their own, thanks to advances in AI and machine learning
Robotic hands could only do what vast teams of engineers programmed them to do. Now they can learn more complex tasks on their own. — A robotic hand?
Context & Ripple Effects
This 2018 piece marks the moment robotic manipulation stopped being a pure programming problem: hands that once did only what engineering teams explicitly coded could now acquire complex tasks through machine learning. The follow-on coverage shows how durable that framing proved — within a year, Google had formalized it in its Robotics at Google program, which applied machine learning to simpler machines than Boston Dynamics' hardware-first approach.
The intervening years tested the claim. A 2020 survey of robotics breakthroughs tracked momentum through the pandemic-era interest surge, while 2024 reporting identified diffusion models, visual language models, and liquid neural networks as the techniques now teaching robots movement and skills. Yet by late 2025, dexterity remained the stated blocker to deploying humanoids in factories and caregiving — meaning the hand, not the brain, is where the learning-vs-programming question gets decided.
First-order effects
- Robotics teams at players like Google can redirect effort from hand-engineering grasp behaviors toward training pipelines, a shift that pits learning-driven programs against Boston Dynamics' demonstration-led hardware tradition.
Second-order effects
- Humanoid robot developers targeting factories and caregiving find their deployment timelines gated by hand dexterity rather than cognition, pushing investment toward tactile sensing and manipulation research.
- AI technique suppliers benefit twice over: the same model families advancing robot skill acquisition are also behind the coding-agent leap Karpathy describes, concentrating talent and compute around learning systems.
Third-order effects
- If learned manipulation keeps improving, the industry's constraint shifts from scarce robotics-engineering labor to data and compute for training — mirroring how software itself is becoming unrecognizable as agents take over coding.
- Dexterity becomes the sorting mechanism for which embodied-AI markets open first: structured factory tasks before unstructured caregiving, since the latter demands finer, more variable motor control.
The trend: Embodied AI is converging on learning-based manipulation as the decisive frontier, with dexterity — not intelligence — determining when humanoids reach real workplaces.