14 reasons robotics is hard

An essay on the blog Secondthoughts.ai, titled after its central claim of 14 reasons robotics is hard, argues that AI progress is visible almost entirely in knowledge work: tasks that happen inside a computer. Physical AI is different. There is no robot equivalent of ChatGPT that ordinary people, or even most of the AI community, can actually try. Instead the public sees demo videos, which the author calls a poor way to judge progress: a clip might show the one success out of a hundred attempts, a scenario staged to dodge unsolved problems, footage edited for apparent speed, or a demo with a suspiciously large number of camera cuts. The World Humanoid Robot Games are noted as a rarer, less cherry-picked public benchmark, with some impressive feats alongside a lot of visible failures.

The essay then works through the specific technical gaps still standing between today's robots and a broadly capable artificial worker. On dexterity: a human hand has roughly two dozen degrees of freedom and about 17,000 tactile sensors, and while some robot hands now reach as many as 27 degrees of freedom, none combine flexibility, sensitivity, strength and reliability the way a human hand does; even fingertips packed with thousands of tactile sensors cannot yet survive heavy use. The author argues the control problem, planning motion to grasp an object, fold a shirt, or flip an omelette, may be as hard as building the hand itself. Computer vision that parses a cluttered scene and figures out where it is safe to step is called unsolved. Planning is framed as needing to scale from a single reach into a cupboard up to cooking a full meal or repairing an engine, while replanning around surprises like a stuck bolt or a child in the kitchen. A separate challenge is context: robots need to know where supplies are kept or how a specific person likes their food, and, unlike large language models, cannot draw on a pool of pre-existing training data anywhere near the size of the text and web data that trained systems like ChatGPT and Claude.

Further chapters cover cooperation (robots, unlike isolated AI agents, will often have to coordinate with people or each other), speed (today's general-purpose robots typically move far slower than humans, which hurts their usefulness as cooks or warehouse workers), strength (powerful motors add heat and drain batteries, and a strong, heavy robot is more dangerous), and mobility (the essay frames the wheels-versus-legs, and how-many-legs, question as still open: wheels are cheaper and more stable, legs handle stairs and clutter). On safety, the author contrasts a self-driving car, which can simply brake or pull over when confused, with a general-purpose robot that might freeze mid-step, drop something, or leave a stove on. On endurance, today's bipedal robots typically run a few hours before recharging, which the author expects workarounds (battery swaps, charging floors, tethers) to solve; the bigger surprise, they write, is that a human-sized robot generates about twice as much heat as a person without the body's own cooling mechanisms, and that frequent breakdowns remain common because the machines are engineered to push size, weight and performance limits. On generalization, the essay points to Waymo, whose cars have driven over 220 million miles, 250 times what a typical American drives in a lifetime, yet have still been observed driving into flooded roads and over burning fireworks; homes and workplaces, it argues, will throw robots far more edge cases than roads do. The essay's text breaks off mid-sentence while introducing a final section on compute, the processing power all of the above will require, so the remaining items in its own 14-item count are not captured here.

Key facts

• A human hand has roughly two dozen degrees of freedom and about 17,000 tactile sensors; some robot hands now reach as many as 27 degrees of freedom, but none match the human hand's combined flexibility, sensitivity, strength and reliability.
• Waymo cars have driven over 220 million miles, 250 times what a typical American drives in a lifetime, yet have still been observed driving into flooded roads and over burning fireworks.
• A human-sized robot generates about twice as much heat as a person and lacks the body's own cooling mechanisms, which the author flags as a bigger challenge than battery endurance.
• Today's bipedal robots typically run for only a few hours before recharging.
• The essay argues demo videos are a poor way to judge robot progress, since a clip can show one success out of many failed attempts, an avoided challenge, edited footage, or heavy camera cuts.

Why it matters

There is a live narrative in AI circles that robots are about to do for physical labor what large language models did for knowledge work, with some forecasts describing the combination as an economic 'explosion.' This essay pushes back with specifics: unlike ChatGPT, there is no widely available physical-AI product a person can test themselves, and the public record is built almost entirely from demo videos, which the author argues systematically overstate readiness through cherry-picking, staging and editing.

Who it affects

Anyone forming an opinion of humanoid robotics from demo footage: investors and journalists covering the sector, engineers and founders building general-purpose robots, and readers trying to judge how close robots are to household or warehouse work. The essay also touches self-driving cars through the Waymo example, since it treats edge-case handling as a shared unsolved problem across physical AI.

How to use it

The author frames the list explicitly as a checklist: next time you watch an impressive robot clip, ask which of dexterity, vision, planning, context-understanding, cooperation, speed, strength, mobility, safety, endurance or edge-case generalization the demo actually shows, and which the scenario might be avoiding. It is a way to read robot marketing more skeptically rather than a product or technique to adopt.

How solid is it

This is an opinion and analysis essay on an independent blog, not a peer-reviewed study; the visible text does not cite sources for its figures, such as the human hand's 17,000 tactile sensors or the 27-degree-of-freedom robot hands, and names no specific companies behind those claims. Its concrete example, Waymo's mileage and edge-case failures, is checkable and specific. The stored text cuts off mid-sentence during the final 'Compute' section, so part of the essay's own 14-point list is not covered here.

Risks and caveats

The piece does not name the companies behind the robot-hand figures it cites, so those numbers should be read as general claims rather than attributed to a specific product. Because the source text is truncated before the essay finishes its compute section, this retelling cannot confirm what the remaining reasons in the '14 reasons' framing are, only the challenges the available text actually covers.

“The control problem may be as challenging as the problem of physical construction.”

— Secondthoughts.ai essay on humanoid robotics