Support CleanTechnica’s work through a Substack subscription or on Stripe.
Humanoid robots have a habit of returning to public attention in waves. Each wave arrives with smoother motion, better balance, and more confident timelines. The claim is usually some version of general purpose capability. The promise is a machine that can safely share space with humans and perform a wide range of everyday tasks in homes and workplaces. This promise has been made repeatedly for decades. Each time, the technical progress is real, but the delivery date moves forward again. The pattern increasingly resembles fusion energy. There is genuine science, steady advances, and real b…
Support CleanTechnica’s work through a Substack subscription or on Stripe.
Humanoid robots have a habit of returning to public attention in waves. Each wave arrives with smoother motion, better balance, and more confident timelines. The claim is usually some version of general purpose capability. The promise is a machine that can safely share space with humans and perform a wide range of everyday tasks in homes and workplaces. This promise has been made repeatedly for decades. Each time, the technical progress is real, but the delivery date moves forward again. The pattern increasingly resembles fusion energy. There is genuine science, steady advances, and real breakthroughs, but the remaining barriers turn out to be deeper than expected, so timelines Doppler into the future.
My own exposure to robotics sits on the periphery rather than at the center of the field. Years ago, I spent time exploring robotics as a side research interest, largely in virtual environments. That work involved simulations of joints, linkages, and kinematics, some algorithmic exploration of control, and a lot of reading across academic robotics literature, including masters and PhD theses from multiple institutions. The long time collaborator, a co-founder of my UK water industry digital twins company Trace Intercept, went further, building physical systems and developing a fully articulated robotic hand as a research project funded by government grants. My interaction with that work was observational and experimental, although focused on the context of commercial use cases, but never advanced to creation of a firm or product. The takeaway from that period was not confidence in rapid progress, but respect for how many layers of complexity sit between a working demo and a reliable system. Even in controlled lab settings, the gap between theory, simulation, and hardware was persistent.
Robotics has advanced significantly, but not evenly. Some problem domains have seen dramatic improvement. Others remain stubborn. Locomotion is the clearest example of progress. Bipedal robots can now walk, run, jump, and recover from disturbances. Quadrupeds can traverse rough terrain. Hybrid wheel leg systems can combine rolling efficiency with stepping over obstacles. This progress is visible in both Western and Chinese research labs and companies. These machines move in ways that would have been extraordinary twenty years ago. Balance, gait planning, and dynamic control are no longer the primary bottlenecks.
What these demonstrations do not show is general usefulness. Movement is a prerequisite for many tasks, but it is not the task itself. Homes and most business environments are dominated by manipulation rather than locomotion. Doors, drawers, tools, dishes, clothing, switches, cables, and fragile objects define the workload. The central challenge is not getting to the object, but interacting with it safely and reliably. This is where progress slows sharply.
Dexterous manipulation remains the central unsolved problem in robotics, and the reasons go well beyond mechanical design. Human hands combine strength, fine force control, dense tactile feedback, compliance, and rapid reflexive adjustment in ways that are deeply integrated and largely subconscious. Robots struggle not because they lack fingers, but because touch itself is a difficult modality to sense, represent, and learn from. Unlike vision and hearing, touch is local and invasive. Information only arrives after contact has begun, at which point the robot is already changing the state of the world.
Compounding this, there is no stable, global representation of touch comparable to pixels or audio waveforms, and no large, standardized libraries of tactile data to learn from, unlike images and audio files. Each tactile sensor has its own geometry, materials, and noise characteristics, making data hard to transfer between systems. Industrial robots avoid these problems by operating in structured environments with known objects, fixed positions, and rigid fixtures, often minimizing contact uncertainty altogether. We can record images of hands doing things, but we can’t record all of the complex things occurring under the skin and in the nerves. Promises of LLMs watching hands do things and recreating control mechanisms for artificial hands are just that, promises, and worth the cost of the digital ink on Powerpoint slides.
Outside those constraints, performance degrades rapidly. Modest variation in object shape, orientation, surface friction, or stiffness can cause grasp failure, slipping, or damage. Adding more fingers or joints does not resolve this. Manipulation requires tightly coupled control, sensing, and learning operating at millisecond timescales, and errors in any layer propagate immediately into physical failure. This is why manipulation remains brittle even as perception and locomotion improve, and why general-purpose robotic hands capable of operating reliably in unstructured human environments remain far from solved. I was deeply impressed with OpenAI’s robotic hand which could solve Rubik’s Cube one-handed, but watching that at the time state of the art effort from 20109 leaves little impression that putting on the end of a robot’s arm was going to be commercialized any time soon. OpenAI shut down that division in any event.
Beyond reliability of operation is the ability to keep doing tasks without breaking down. A core reason general-purpose humanoid robots remain elusive is that fine-grained manipulation with human-like dexterity is inherently failure-prone. This is not simply a matter of better software or more data. Human hands are biologically redundant, compliant, and deeply fault-tolerant systems. They combine dozens of degrees of freedom with soft tissue, dense tactile sensing, and rapid reflex loops that absorb error rather than amplify it. Small slips, misjudgments, or unexpected contacts are corrected continuously and subconsciously. When humans fail at manipulation, failure is usually graceful.
Mechanical systems do not share these properties. Robotic hands are rigid, sparsely sensed, tightly constrained, and intolerant of deviation. As dexterity increases, so do the number of joints, sensors, control loops, and timing dependencies, and with them the number of ways the system can fail. Capability does not scale linearly. Failure modes multiply faster than performance improves.
These technical challenges translate directly into mechanical reliability problems. Highly dexterous manipulators require many small actuators, transmissions, sensors, and cable runs, all operating under tight tolerances and frequent load reversals. Each additional degree of freedom introduces wear points, backlash, calibration drift, and failure risk. Unlike simple industrial grippers, sophisticated hands are exposed to impacts, misaligned forces, contamination, and thermal variation during normal operation. A big part of my modeling and research a couple of decades ago was spent on joints, and then spent eliminating joints, looking for simple physical intelligence approaches that could replace them entirely.
Mean time between failures (MTBF) tends to fall as mechanical complexity rises, not improve. Servicing such manipulators requires skilled labor, replacement parts, and downtime, all of which undermine economic viability. For robots intended to operate continuously in homes or businesses, low MTBF is unacceptable. Even infrequent mechanical failures become costly and disruptive at scale, reinforcing why increasingly sophisticated manipulators struggle to transition from lab demonstrations to reliable, maintainable products in real-world environments. Certainly this was my collaborator’s experience with his mechanical hand, which failed constantly even in its relatively simple form and without repeatedly doing any tasks beyond moving.
Rodney Brooks is one of the foundational figures in modern robotics, with decades of academic and industrial work that shaped how the field thinks about embodied intelligence. His research on subsumption architectures fundamentally challenged model heavy, top down approaches and instead emphasized behavior, interaction, and robustness in real environments. His work was among the core bodies of research that my collaborator and I studied closely during our own exploratory work in robotics, and its influence carried forward into multiple other domains we explored or worked in professionally afterward. Brooks later co founded iRobot, best known for the Roomba, and helped translate those ideas into one of the few genuinely successful consumer robots, giving his views unusual weight across both research and commercialization.
Brooks has been making these arguments since his early work on industrial robotics and human worker safety, well before collaborative robots were commercially viable. His research was directionally correct but arrived before the actuators, sensors, and control systems needed to make safe human robot interaction practical. One of his core points was that motor driven arms and legs carry significant kinetic energy, and when a system becomes unbalanced or control degrades, that energy is released in unpredictable ways. A robot strong enough to be useful can easily injure a nearby human if a joint overshoots, a perception system misclassifies an obstacle, or a control loop destabilizes.
Brooks consistently emphasized that reliability has to be extraordinarily high, not merely acceptable most of the time, because rare failures become unavoidable at scale. Industrial robots addressed this by keeping people and machines apart or by tightly constraining motion and force. Collaborative robots only work because they are intentionally slow, weak, and limited. A general purpose humanoid combines high force, long reach, autonomy, and constant proximity to people, which makes certification, insurance, and large scale deployment far more difficult than the demonstrations suggest.
This is why narrow robots have succeeded where humanoids have not. Robotic vacuum cleaners perform a single task in a constrained environment. Autonomous electric lawn mowers do the same outdoors, and indeed my collaborator and I were working on exactly that use case as a core exploration, just a couple of decades too soon. Warehouse robots move standardized goods along known paths. These systems are commercially viable because their scope is limited. Safety envelopes are manageable. Failure modes are predictable. Economics are clear. They are not early versions of humanoid robots. They are a different branch of the robotics tree, optimized for value rather than generality.
Boston Dynamics provides a useful reality check. The company has led the world in legged locomotion, with Atlas demonstrating balance, agility, and complex movement that few others can match. Despite this, Atlas remains a research platform rather than a commercial product. The company has focused its business on non humanoid systems such as Spot for inspection and Stretch for warehouse material handling, both designed for constrained tasks and environments.
Financial disclosures and transaction filings across successive owners show a consistent pattern of large losses. Under Google, Boston Dynamics operated as a research focused organization with substantial annual burn and little revenue. Under SoftBank, reporting around the acquisition and sale indicated continued operating losses, adding several hundred million dollars cumulatively. Since acquisition by Hyundai, losses reported within Hyundai’s robotics segment have remained significant, roughly $100 to $200 million per year, even as products reached the market. Taken together, a reasonable reading of available filings and disclosures suggests that Boston Dynamics has likely accumulated total losses well in excess of $1 billion over the past two decades. After decades of technical leadership and sustained backing from some of the largest technology and industrial firms in the world, general purpose humanoid robots have still not become an economic proposition. When Google, Honda and hundreds of robotics engineers working for two decades can’t crack the problem, expecting it to be cracked on venture capitalist return timeframes is deeply unrealistic.
The Roomba robot vacuum cleaner story reinforces the same lesson from the opposite direction. Rodney Brooks helped shape Roomba by deliberately avoiding general intelligence and humanoid form. The robot solved one narrow problem using simple sensors and behavior based control, a subsumption architecture. It became one of the most successful consumer robots ever sold. Even so, the business struggled once competition increased and margins collapsed. After a failed acquisition attempt, the parent company entered bankruptcy proceedings in 2024. (Brooks hadn’t been involved with the company for a long time, having left his Board chairman role in 2016 and forming another firm.) This was the easy case in robotics. A simple machine, a clear value proposition, and still a hard business. Scaling from that experience to a general purpose humanoid is not a linear step.
China’s robotics ecosystem often features prominently in discussions of humanoids. Promotional videos show impressive motion, including martial arts routines, acrobatics, and hybrid wheel leg machines that move quickly over complex terrain. I’ve been fascinated watching key robots advance in YouTube videos in the past two years. These demonstrations are real and technically competent. What stands out is what they lack. Manipulators are basic. Hands are coarse. Grippers are rigid and limited. In industrial settings, some Chinese firms deploy vaguely humanoid robots with dual arms on wheeled bases and a sensor cluster mounted where a head would be. CATL provides a clear example. The robot looks humanoid in that it has a torso, a head and two arms, but the form factor being vaguely human, if you squint, adds little functional value. The manipulators are crude, with limited dexterity and force control, being designed to pick up square edged components and slide them into specific spaces. They are little more capable than the mechanical arms sketched decades ago, with hands that look a lot more like Rosie the Robot’s from the Jetsons than to a human worker. It has an omnidirectional wheeled platform suitable for polished factory floors and nothing else, and a single hinged ‘leg’ that allows it to adjust its height a bit. The humanoid appearance is limited and more marketing than reality. The capability remains narrow and task specific.
Tesla’s Optimus program does not justify a material valuation premium at this stage, despite market narratives to the contrary. Depending on assumptions, analysts and investors appear to be implicitly assigning tens of billions of dollars, and in some cases over $100 billion, of Tesla’s market capitalization to humanoid robotics optionality. That valuation is not supported by demonstrated capability, timelines, or commercial readiness. Public Optimus demonstrations have repeatedly shown signs of immaturity, including periods where the robot was teleoperated rather than autonomous, limited manipulation that struggled with simple objects, and staged environments designed to avoid unpredictable interaction. In some cases, demonstrations required human intervention or reset when tasks failed, underscoring how brittle the system remains. These are normal and expected outcomes for an early research platform, but they are inconsistent with claims of near-term general-purpose deployment.
A commercially available humanoid robot that can safely operate around humans, perform diverse tasks without supervision, and meet certification and liability requirements is likely decades away. Assigning a large valuation premium to Optimus today assumes breakthroughs in manipulation, safety, autonomy, and integration that have not yet occurred and won’t arrive on investor-relevant timelines. It’s just another part of Tesla’s meme stock status, along with full self driving which gets halfway there with every step, but as a result never arrives, and its claims of AI leadership.
The fusion analogy holds because it captures both optimism and restraint. Fusion energy is not a fraud, per se. Progress is real. The same is true for humanoid robotics. Each generation solves some problems and reveals others, it’s just that the next problems are often much harder than the problems just solved. Fusion claims about maintaining a very hot plasma for a little longer are indeed breakthroughs, but are revealing incredibly complex things that hadn’t been much more than guessed and which require more decades of engineering. Integration turns out to be harder than expected. Safety and economics dominate late in the process. Timelines move forward by years or decades again. In the meantime, robotics continues to succeed where constraints are clear and goals are narrow. The future is likely to be filled with many specialized robots quietly doing useful work, not humanoids walking through kitchens and offices. The science and engineering are advancing. The destination is just much farther away than the demos suggest.
Sign up for CleanTechnica’s Weekly Substack for Zach and Scott’s in-depth analyses and high level summaries, sign up for our daily newsletter, and follow us on Google News!
Advertisement
Have a tip for CleanTechnica? Want to advertise? Want to suggest a guest for our CleanTech Talk podcast? Contact us here.
Sign up for our daily newsletter for 15 new cleantech stories a day. Or sign up for our weekly one on top stories of the week if daily is too frequent.
CleanTechnica uses affiliate links. See our policy here.
CleanTechnica’s Comment Policy