SWITCHED ON
The daily technology series nobody asked for but everyone needed
The Humanoid Race: Figure, Optimus, and Robots Built for Human Spaces
The pitch is simple: build a robot shaped like a person, and it can work in the millions of spaces already built for people, without redesigning a single warehouse, kitchen, or factory floor. Whether that pitch survives contact with the actual difficulty of bipedal manipulation is the question this episode tries to answer honestly.
Every previous wave of industrial robotics succeeded by redesigning the task for the robot — a fixed arm on an assembly line, a purpose-built warehouse picking system. Humanoid robotics makes the opposite bet: redesign the robot for the task humans already do, in the environments humans already built. It is a more ambitious bet and a more difficult engineering problem than almost any previous robotics generation has attempted.
Yesterday we opened Season Three with agentic AI — the shift from conversational chatbots to autonomous task-completing agents, computer use as the browser-native version of that shift, coding agents as the most mature deployment, and the verification challenges that autonomy introduces. Today we are taking that same underlying AI capability and asking what happens when it gets a body. Humanoid robotics: the increasingly serious and increasingly well-funded race to build general-purpose robots shaped like humans, capable of working in the same physical spaces humans already occupy, and the genuine engineering difficulty that stands between the demo videos and a robot doing useful, reliable work in your warehouse, kitchen, or home.
01 — Why Humanoid Form Factor at All
The case for building robots in a roughly human shape — two arms, two legs, a head-height sensor package, human-scale proportions — is fundamentally an economic and infrastructural argument rather than a purely technical one. The built environment of factories, warehouses, kitchens, offices, and homes has been designed around human dimensions and human capabilities for centuries: door heights, shelf reaches, stair dimensions, tool handle sizes, vehicle cabin layouts. A robot that matches human form factor can, in principle, operate in this environment without requiring it to be redesigned or retrofitted, which is an enormous practical advantage relative to specialised robots that require purpose-built environments.
The counter-argument, which serious roboticists raise consistently, is that the human form is a remarkably difficult engineering target precisely because evolution optimised it for a general-purpose combination of mobility, dexterity, and energy efficiency that took hundreds of millions of years to arrive at, and that current robotics technology is nowhere close to matching in several critical dimensions simultaneously. A wheeled robot is more stable and energy-efficient than a bipedal one for most warehouse floor environments. A fixed robotic arm is more precise and more reliable than a humanoid arm for most manufacturing tasks. The humanoid bet is not that the human form is optimal for any single task — it demonstrably is not — but that its generality, its ability to use the same tools and navigate the same spaces humans do without modification, provides enough economic value to justify the additional engineering difficulty.
Bipedal walking, which humans learn by around eighteen months of age with what feels like no conscious effort, is genuinely one of the hardest control problems in robotics — a continuous balancing act across an inherently unstable system, solved by biological neural systems refined over evolutionary timescales that robotics engineers are still working to approximate computationally.
02 — Who Is Actually Building These Things
The current humanoid robotics landscape includes a genuinely diverse set of approaches and funding models. Figure AI, backed by significant investment from OpenAI, Microsoft, NVIDIA, and Jeff Bezos among others, has deployed its Figure 02 robot in limited commercial trials with BMW's manufacturing operations, performing specific structured tasks including sheet metal part insertion — a deliberately narrow, well-specified initial application chosen precisely because it plays to current humanoid robotics' strengths in controlled, repetitive environments rather than the more ambitious general-purpose claims that marketing materials sometimes imply.
Tesla's Optimus programme, which has received the most public attention of any humanoid effort given Elon Musk's public profile and Tesla's existing manufacturing and AI infrastructure, has demonstrated increasingly sophisticated capability in choreographed demonstrations, though independent, unscripted evaluation of its actual autonomous capability — as opposed to teleoperated demonstrations, where a human operator remotely controls some or all of the robot's actions during a "demo" — has been a persistent point of scepticism among robotics researchers, some of whom have specifically questioned how much of Tesla's public Optimus demonstrations reflected genuine autonomous capability versus human teleoperation. This distinction matters enormously for assessing actual technological progress versus marketing.
Boston Dynamics, the longest-established name in advanced robotics and now owned by Hyundai, has taken a notably different public communication strategy — its Atlas humanoid robot has demonstrated genuinely startling dynamic mobility (backflips, complex parkour-style movement) in research demonstrations for years, reflecting the company's deep engineering heritage in dynamic robotic locomotion, while being comparatively more conservative and evidence-based in its claims about near-term commercial deployment timelines, a communication approach that reflects the company's longer history of navigating the gap between research demonstration and reliable commercial product.
03 — The Manipulation Problem
Walking, despite being genuinely difficult, has actually advanced faster in recent years than the arguably harder problem of dexterous manipulation — using hands to grasp, hold, and manipulate objects of varying size, shape, weight, and fragility with the reliability and adaptability that human hands demonstrate almost unconsciously. This is, if anything, the more consequential bottleneck for humanoid robots achieving genuine economic usefulness, because most valuable physical labour tasks — from warehouse picking to cooking to caregiving to most manufacturing assembly — depend far more on sophisticated manipulation than on locomotion.
The human hand contains twenty-seven bones, dozens of muscles and tendons, and an extraordinarily dense concentration of tactile sensory receptors that provide continuous feedback about grip force, texture, temperature, and slippage — feedback that the brain processes largely unconsciously to make constant micro-adjustments during even simple grasping tasks. Current robotic hands, even the most advanced research-grade versions, remain meaningfully less capable across this full range of sensing and adjustment, particularly for tasks involving deformable objects (fabric, food items, packaging materials of varying rigidity), objects requiring precise force calibration (an egg versus a brick), or tasks requiring the kind of improvised, in-the-moment adaptation that humans perform without conscious deliberation when a grip starts to slip or an object turns out to be unexpectedly heavy.
04 — The AI Model Behind the Body
The renewed wave of humanoid robotics investment, distinct from previous robotics cycles, is driven substantially by advances in the AI models controlling these robots rather than purely by mechanical engineering improvements. Vision-language-action models — AI systems trained to process visual input and natural language instructions and output physical robot actions, extending the same transformer architecture underlying large language models into the domain of physical control — represent the most significant recent advance, allowing robots to be instructed in natural language ("pick up the red cup and place it on the shelf") and to generalise that instruction across variations in object position, lighting, and environment in ways that earlier hand-programmed robotic control systems could not.
Google DeepMind's RT-2 and subsequent models, along with similar systems from Physical Intelligence and other robotics AI labs, have demonstrated that training robotic control policies on internet-scale data (combining robot demonstration data with general web image and text data) produces meaningfully better generalisation to novel objects and situations than training on robot-specific data alone — an approach that mirrors, in the physical domain, the scaling insights that drove large language model progress. This is a genuinely significant technical development, and it is the primary reason serious robotics researchers, who have watched previous humanoid robotics waves overpromise and underdeliver across multiple prior decades, are taking the current wave more seriously than its predecessors.
05 — The Realistic Timeline and Where the Value Actually Lands First
The honest assessment of humanoid robotics' near-term trajectory, consistent with the pattern this series has documented across dozens of technologies, is that meaningful commercial deployment is beginning now in narrow, well-controlled applications — Figure's automotive manufacturing trials, Agility Robotics' Digit humanoid performing structured warehouse tasks for Amazon and other logistics operators — rather than the general-purpose, work-anywhere robots that the most ambitious marketing narratives project.
The economics currently favour deployment in structured industrial and logistics environments over homes or unstructured public spaces, for the straightforward reason that industrial environments can be somewhat adapted to accommodate current robot limitations (consistent lighting, predictable object types, defined work areas), while homes and public spaces present the full, uncontrolled variability of the real world that remains genuinely difficult for current manipulation and perception capability to handle reliably. The most credible timelines from serious industry participants — as opposed to the most optimistic public statements from company founders with obvious incentives to generate investor and public excitement — suggest narrow industrial and logistics deployment expanding meaningfully over the next three to five years, with genuinely general-purpose, reliable home and unstructured-environment deployment remaining a longer-horizon prospect, likely a decade or more out, contingent on manipulation and perception capabilities that have not yet been demonstrated at the reliability required for unsupervised operation around untrained members of the public.
Tomorrow we are going somewhere that connects directly to the agentic AI and robotics threads opening this season — AI safety evaluations and red-teaming, the increasingly formalised practice of stress-testing AI systems before release, and what the emerging science of finding an AI model's failure modes actually looks like in practice. See you then.
Switched On is a daily technology series covering the ideas, systems, and arguments shaping the digital world. Opinionated. Witty. Occasionally wrong. Always worth the argument.
