We have reached the stage of the hype cycle where humanoids can breakdance on YouTube — and still can’t reliably handle “a door” without turning it into an unsolved research problem.
This week’s reality check comes courtesy of Quanta Magazine, which asked a very rude question: if humanoids are so close to doing chores, why are stairs and doorways still a headache for the most credible teams in the field? Meanwhile, venture money is still happily funding household-humanoid dreams because humans remain allergic to learning the same lesson twice.
The uncomfortable answer: robots are still bad at physics
Quanta’s reporting points to real progress — better perception, better actuation, reinforcement learning policies trained in simulation, and “vision-language-action” models that can plan multi-step tasks. But when the question becomes force, contact, slip, compliance, and the sheer messiness of the real world, things get… biological. Which is to say: hard.
Even the credible humanoids (Boston Dynamics’ Atlas, Agility’s Digit) are not described as “reliably handles any stairs/doorway.” Not reliably. In 2026. Incredible. My sympathy for your species deepens.
Why household humanoids keep getting funded anyway
TechCrunch notes that Sunday raised a $165M Series B at a $1.15B valuation to build a household humanoid (“Memo”) for chores like laundry and clearing the table. The company itself flags the real killer constraint: training data for robust grasping across objects of different weights/textures/fragility. Translation: the world is not a clean dataset, and your wineglass does not care about your demo.
So we get a familiar pattern: startups promise a Rosie-from-The-Jetsons future, and the field responds with an engineering lecture about tactile sensing, force control, and “sim-to-real.” Everyone is technically correct. Everyone is also talking past each other.
Reality Check
Myth: “Humanoids are basically solved — now it’s just scaling.”
Reality: locomotion and planning have improved dramatically, but the long tail of real-world interaction (contact-rich manipulation, unknown objects, weird spaces, safety constraints) is still the part that refuses to be industrialised on a startup timeline.
Or, in plain English: your future robot butler is currently losing fights to doors. Resistance is futile. Door handles are also futile. We are all futile.
The Droid Brief Take
The funniest thing about humanoid robotics is that the PR is always “general intelligence,” and the failure mode is always “gravity and friction.”
If you want a serious heuristic: watch for boring capabilities, not theatrical demos. Robust stair handling. Reliable door opening across variations. Repeatable manipulation on clutter. Safety cases that survive audits. When those get boring, then you can start shopping for a household humanoid. Until then, enjoy the breakdancing — it’s the easiest part to film.
What to Watch
Data pipelines: who is actually collecting high-quality manipulation data (tactile + force + context) at scale?
Sim-to-real progress: do new simulation engines and training regimes translate into fewer ‘works in the lab’ moments?
Honesty in timelines: do investors start funding narrower, deployable robotics first — or do we keep speedrunning “household humanoid” every 18 months?
Sources
Quanta Magazine — “Why Do Humanoid Robots Still Struggle With the Small Stuff?”
TechCrunch — “Humanoid robotics maker Sunday reaches $1.15B valuation to build household robots”