Artificial intelligence has learned to think. Teaching it to move, sense and act in our physical world is the harder frontier — and the timeline to an Asimov-style future depends on hardware and trust, not on smarter algorithms.
A conversation published this month by ChinaTalk with two robotics analysts from SemiAnalysis made a claim worth sitting with: general-purpose robots may be the first technology in a century to genuinely decouple physical production from human labour. The exchange centered on Unitree’s rapid ascent and China’s grip on actuators and rare-earth processing, but the strategic question it raises is broader than any single company. If cognition has already been industrialized through large language models, what is actually standing between us and a world in which machines share our physical spaces as capably as they now share our inboxes — an Asimov-like future, if you will, of robots that cook, care, build and coexist? The honest answer is that the constraint has shifted. It is no longer whether a model can plan a task; it is whether a body can execute it reliably, safely and cheaply enough to matter.
The Chasm Between Demo and Deployment
McKinsey’s industrials practice, tracking humanoid pilots across manufacturing and logistics, frames the situation as a chasm rather than a slope: pilots at BMW, Mercedes and Amazon already show robots handling structured, mapped tasks such as intra-factory logistics, but scaling beyond those controlled settings requires closing four separate gaps at once — technical performance, cost, safety assurance and regulation. None of these gaps yields to better language models alone. A humanoid still needs billions of simulated or teleoperated interactions to master even narrow manipulation tasks, and it generalizes poorly once the environment changes. Tactile sensing, adaptive hands and actuators dense enough to fit inside a human-scale limb remain, per McKinsey’s own account, unsolved engineering problems rather than data problems. Progress on the cognitive side has simply outpaced progress on the body that has to carry it out.
Four Roadblocks, Not One
The first roadblock is dexterity and endurance: hands, joints and cooling systems that can work an eight-hour shift without overheating or degrading, a threshold today’s platforms approach only in narrow, task-specific configurations. The second is data: embodied experience cannot be scraped from the internet the way text can, so every gain in fine manipulation currently depends on custom rigs, teleoperation fleets or purpose-built simulators, an expensive and slow substitute for the vast corpora that trained today’s language models. The third is the supply chain itself. Georgetown’s Center for Security and Emerging Technology notes that scalable humanoid manufacturing depends on cost curves that have not yet bent, and that the actuator and rare-earth ecosystem underpinning them is geographically concentrated to a degree that should worry any executive planning a five-year deployment. The fourth, and least discussed in engineering circles, is trust: collision avoidance, cybersecurity and decision transparency have to be certified before a machine with real torque is allowed to work beside an unprotected human, and that regulatory scaffolding barely exists yet.
Closing the Gap: Three Plausible Levers
None of these roadblocks is unsolvable, and the direction of travel is already visible. Vertical integration is the first lever: companies that control actuator design, from the rare-earth magnet to the finished joint, iterate faster because every failure feeds directly back into the next revision, the same feedback loop that let China’s electric-vehicle industry overtake incumbents once dismissed as unthreatening. The second lever is a deliberate, jagged deployment strategy rather than a search for one universal robot: pairing a narrower, well-instrumented task — cabling in a data center, cabin cleaning at an airport, sheet-metal handling on an assembly line — with custom end effectors and bespoke models closes the dexterity gap far faster than waiting for a general-purpose foundation model to arrive. The third lever is allied and diversified sourcing: economic zones with faster permitting, partnerships across Japan, Australia and Southeast Asia for rare-earth processing, and industrial policy that treats actuators the way semiconductors were treated a decade ago. None of this is glamorous, but neither was the decade of unglamorous capital that built the modern chip industry.
A Speculative Timeline
Based on where these levers currently stand, a defensible speculation looks something like this. Within two to four years, coarse manipulation on mobile platforms becomes deployable at meaningful scale across logistics and manufacturing, essentially extending today’s pilots into standard procurement lines. By the end of the decade, a second wave of tasks involving genuine tactile judgement, from wiring to light assembly, starts crossing from pilot to platform as sensing and actuator costs fall by the five-to-tenfold factor that McKinsey estimates is still required. The household robot — the one that folds laundry, cooks and keeps an eye on an elderly relative — sits on a longer curve, not only because the manipulation problem is categorically harder, but also because unsupervised proximity to untrained humans multiplies the safety and liability burden well beyond anything a factory floor demands. A realistic Asimov-like coexistence, where physical AI is simply part of how homes and cities function, is plausibly a 2035-to-2045 story rather than a 2030 one, and it will arrive unevenly: industrial and logistics environments first, controlled public settings such as hospitals and airports second, private homes last.
What This Means for Decision-Makers Today
For the executives reading this, and the reference here is mainly for weastern people, the practical takeaway is not to wait for certainty before moving. The organisations already piloting humanoids in narrow, well-defined tasks are the ones building the operational muscle, such as data pipelines, safety protocols, supplier relationships, that will compound once the hardware curve bends. Physical AI is following the trajectory digital AI already showed us: progress arrives unevenly, adoption lags capability, and the businesses that treat this as a five-year infrastructure question, rather than a single dramatic product launch, are the ones that will be ready when the robots stop dancing and start working.
