Watch the viral videos and it is tempting to conclude the robot revolution is already here: humanoids vaulting, folding shirts and loading washing machines. Yet beneath the spectacle a growing chorus of engineers and researchers says the problem is not the software but the bodies those brains are trying to animate. Sony’s research programme, for example, has explicitly flagged the “limited number of joints” in today’s humanoids and is calling for work on “flexible structural mechanisms” to enable genuinely dynamic motion. According to that call, smarter, more adaptive physical architectures are needed before the machines can realise their promise. (Sony’s funding initiative invites universities and labs to propose cross‑disciplinary projects aimed at those exact shortcomings.) [1][2]

At the heart of the debate is a familiar trade‑off: robots built around a “brain‑first” model must rely on heavy computation and powerful actuators to correct for bodies that do not help themselves. That approach is inefficient. Recent coverage singled out figures that capture the scale of the mismatch — noting, for example, that one commercial prototype consumes roughly 500 watts just to walk, while a human on a brisk walk uses around 310 watts — and argued that rigid, sensor‑poor limbs force machines to expend energy to overcome their own inertia. Theoretical work on morphological computation explains why this matters: bodies can, by design, take on part of the control problem, stabilising motion and pre‑processing physical interactions so controllers need do less work. [1][4]

The result is diminishing returns from ever‑bigger neural models and more elaborate vision stacks. High‑profile demonstrations have themselves become a focal point for sceptics. When Tesla shared footage of Optimus folding a shirt, commentators quickly pointed out staged elements and gaps in autonomy; Business Insider reported that Elon Musk later acknowledged the action was “not yet” autonomous and that demonstrations can involve teleoperation or staged assistance. Such episodes underline the gap between polished showreels and robust, general‑purpose manipulation in messy, unpredictable environments. Industry observers warn that impressive demos do not yet equate to dependable real‑world capability. [1][6]

Nor do the acrobatic reels tell the whole story for established research platforms. Boston Dynamics’ all‑electric Atlas has been redesigned to trade hydraulics for electric actuation and showcases an extraordinary range of motion, but even the company’s own evolution illustrates the point: Atlas remains a research prototype rather than a commercial worker. Reporting on the April 2024 reveal stressed that the viral routines mask continuing limitations — from energy and autonomy to the lack of tactile, conforming feet — that make confident, everyday operation on uneven, natural terrain elusive. [1][7]

That practical shortfall is precisely what the emerging field of mechanical intelligence (MI) is trying to address. Researchers at London South Bank University’s MI Research Group, led by Hamed Rajabi, translate biomechanics and adaptive‑structure ideas into compliant mechanisms, soft–rigid hybrids and energy‑storing elements that embed “passive” intelligence into the body itself. The principle is old and well documented in robotics literature: morphological computation shows how structures — a pine cone’s humidity‑driven scales, a hare’s elastic tendons, a human fingertip’s soft flesh — can perform useful, automatic responses to the world without expensive sensing or control. Rajabi and colleagues argue that incorporating those lessons into humanoid design would let the controller do higher‑level tasks while the body handles local adaptation. [1][3][4]

There are already practical precedents for the gains MI promises. MIT’s legged‑robot work — exemplified by the Cheetah project — demonstrates how tendon‑like elements, low‑inertia limbs and energy‑regenerative control can yield a cost of transport comparable to animals and markedly extended operating time. Other academic teams are producing hybrid hinges and compliant joints that combine precision with shock absorption, which could let shoulders, knees and hands move with multiple degrees of freedom without constant active correction. These projects suggest that, when mechanical design and control are co‑optimised, robots can become both more capable and more energy‑efficient. [5][3][1]

So why haven’t the industry giants pivoted hard towards MI? Partly because today’s market leaders are rooted in software and electronics ecosystems that favour high‑precision motors, sensors and processors — the very components a brain‑centric strategy exploits. Building mechanically intelligent bodies at scale demands different materials, new manufacturing chains and cross‑disciplinary partnerships that the sector’s supply networks are not yet structured to deliver. Sony’s awards programme and similar initiatives are attempts to bridge that gap by funding collaborations between manufacturers, materials scientists and roboticists to move compliant, adaptive structures from lab curiosity to industrial reality. [1][2]

The implication for the next phase of robotics is clear: progress will be less about choosing hardware over software and more about making them co‑equal partners. As Hamed Rajabi wrote in his analysis, the pathway out of the current humanoid trap involves bodies that shoulder some of the computation — leaving AI to focus on strategy, learning and complex interactions. Researchers and funders are already taking tentative steps in that direction, but turning prototypes and academic proofs into reliable, mass‑manufacturable systems remains the enduring engineering and industrial challenge. Until that synthesis is achieved, the epochal promise of humanoid robots stepping fully into everyday life will remain a work in progress. [1][3][4][2]

📌 Reference Map:

Reference Map:

Source: Noah Wire Services