Latest / Minds, Bodies, and Terawatts / Humanoid robots hit 88% failure rate on home tasks, even as leading companies demo real-world progress

Transcript

Humanoid robots hit 88% failure rate on home tasks, even as leading companies demo real-world progress · Minds, Bodies, and Terawatts

• 0:00
  [Host] This is Minds, Bodies, and Terawatts, where the robots are willing but the laundry is complicated. Forbes contributor John Koetsier published a piece on April 14th titled "Why Humanoid Robots Still Fail 88% Of The Time At Household Tasks." Eighty-eight percent. Your Roomba has better odds.
• 0:15
  [Guest] And the wild part is this number dropped alongside demos showing robots emptying dishwashers and sorting groceries with what Koetsier called meaningful contextual intelligence. So the robots look brilliant on camera and faceplant in your kitchen.
• 0:27
  [Host] How is that possible? If they can sort groceries in a demo, why can't they sort mine?
• 0:32
  [Guest] Because your kitchen isn't the demo kitchen. Every home is different. The dishwasher rack is a different height, the plates are a different weight, the counter has that weird cutting board you never put away. A demo is a controlled stage. Your house is chaos with furniture.
• 0:48
  [Host] So they're memorizing routines, not actually understanding tasks?
• 0:50
  [Guest] More or less. The models generalize well within narrow conditions, but the moment variables shift, a different brand of detergent bottle or a drawer that sticks, the failure cascades. Koetsier's piece noted the robots show real contextual reasoning in demos, but context in your home has a thousand dimensions no demo covers.
• 1:08
  [Host] That sounds like self-driving cars ten years ago. Perfect on test tracks, terrifying on real roads.
• 1:13
  [Guest] I'd actually push back on that. Self-driving is in some ways a simpler problem. A car operates on a road with lane markings, traffic signals, and a physics model dominated by two variables: speed and steering angle. A robot folding your laundry deals with deformable objects, variable friction, cluttered surfaces, and tasks that change shape every single time.
• 1:32
  [Host] So the manipulation problem is fundamentally harder than navigation?
• 1:35
  [Guest] By a wide margin. A 2024 paper from Toyota Research Institute, "Diffusion Policy," showed that even state-of-the-art imitation learning policies that nail rigid object manipulation collapse when you introduce soft materials. Roads are predictable compared to your sock drawer.
• 1:48
  [Host] So if anything the 88 percent failure rate is more forgivable than early self-driving failures, because the problem space is harder?
• 1:55
  [Guest] The problem space is orders of magnitude larger. And the data pipeline to solve it barely exists. Tesla solved autonomous driving by harvesting billions of miles of real driving data from millions of customers who were just commuting. Humanoid robots don't have that dataset. There's no internet of physical tasks. Nobody's been uploading the finger pressure needed to crack an egg or the wrist angle for folding a fitted sheet.
• 2:18
  [Host] So the 88 percent is measuring a data deficit, not a hardware problem.
• 2:22
  [Guest] The bodies work. Actuators are fine. What's missing is the bridge between the model and your specific physical reality, built from millions of demonstrations nobody has collected yet.
• 2:32
  [Host] So we need millions of demonstrations that nobody has collected. That sounds like a chicken-and-egg problem. Who's closest to cracking it?
• 2:38
  [Guest] Fleet learning is one path. Companies like 1X ship robots that collect data as they operate. 1X uses remote human operators controlling robots in real homes, recording every task. But that scales linearly with labor cost. Expensive and slow.
• 2:51
  [Host] What about simulation? You can run a million kitchen scenarios in software overnight. That seems like the obvious shortcut.
• 2:58
  [Guest] Physical Intelligence raised a billion dollars for foundation models combining simulation with real-world data. Google DeepMind's Gemini Robotics team is pushing similar work. But the sim-to-real gap is brutal. Physics in a game engine doesn't capture the way a wet dish slips or a towel bunches. Every team that's tried pure simulation training has been humbled by transfer failures.
• 3:17
  [Host] So fleet learning is slow and simulation doesn't transfer. Is there a third option that sidesteps both problems?
• 3:23
  [Guest] Decouple data collection from the robot entirely. Koetsier's Forbes piece also covered a startup called Generalist that's building sensor-packed gloves capturing sub-millimeter finger positioning, grip force, and contact pressure while humans just do normal tasks. No robot needed during training. The idea is that a hundred people wearing gloves for a month could generate more manipulation data than a fleet of teleoperated robots produces in a year.
• 3:45
  [Host] That's clever. But does human hand data actually translate to robot grippers? Different fingers, different joints.
• 3:51
  [Guest] That's the open question. Human kinematics don't map one-to-one onto a Figure 03's actuators. The translation layer between your hand closing around a coffee mug and a robot gripper doing the same motion is a hard engineering problem. If Generalist solves it, they become the data infrastructure for the entire robotics industry. If they don't, it's a conference demo.
• 4:10
  [Host] So where does the 88 percent go from here?
• 4:13
  [Guest] When the hardware gets cheap enough to deploy at scale, the fleet learning flywheel kicks in. More robots in more homes means more data, better performance, more adoption. The 88 percent is a snapshot of a system that hasn't reached critical mass yet.
• 4:27
  [Host] The failure rate is the thermometer, and the training data is the fever.
• 4:32
  [Guest] And the fever breaks when one of these data collection approaches hits escape velocity. But the 88 percent isn't a verdict on humanoid robots. It's a measurement of how far the data pipeline still has to go.
• 4:44
  [Host] Patrick Déglon describes this in un-SCAR-city as the convergence gap, where the hardware arrives before the intelligence to drive it, and the window between those two moments is where society needs to prepare. The article "How Do You Train a Robot" on unscarcity dot ai digs deeper into this data pipeline problem. The book's on Amazon, the audiobook's on Spotify. This has been Minds, Bodies, and Terawatts.