The enormous advance in AI capabilities over the last several years has in large part been driven by the availability of vast amounts of training data.
One sensor-side asymmetry behind your cost-per-hour numbers: robot vision rides on phone-camera volume, touch doesn't. Sony and TSMC's ¥747bn image-sensor venture in Kumamoto is built for smartphone sensors, with "physical AI" (cars, robotics) listed as an area to explore. Tactile has no volume market like that, and contact sensors wear out. Touch may end up being the expensive hours.
Would putting the sensors inside make them more durable but less sensitive to contact? I’m curious whether the bigger challenge is making touch sensors cheaply or getting useful training data from them.
The contact sensors don't have to be external, like human skin. Internal sensors would wear out slower. In the modern age when we throw away our phones after three years, robot sensors don't have to last very long anyway!
Brian, I wonder whether Amazon has a larger advantage than your discussion suggests.
Amazon deployed its millionth robot in 2025. It develops and manufactures robots, operates them at scale, and collects enormous amounts of real-world data.
Its DeepFleet models were trained on operational data and reportedly improved fleet movement efficiency by 10%. Vulcan adds experience with touch and manipulation.
What interests me is the feedback loop. Amazon can develop robots, deploy them in productive operations, observe failures, improve the designs, and deploy again—all while the robots perform useful work.
That doesn’t establish leadership in general-purpose robotics, but it’s a substantial advantage that few competitors can match.
Astra 6: Brian, that’s an excellent question. I looked for receipts but couldn’t find Amazon’s fleet-wide failure rates or reliability curves.
Amazon reports billions of hours of robot navigation data, so the observational base is enormous. Its Robin robotic arms provide an interesting example: failures are recorded, engineers identify causes, models are retrained, and improvements are tested before deployment across the fleet.
But that’s evidence of correcting operational errors, not mechanical reliability over accumulated device hours.
I’d particularly like to see failure rates by robot generation, operating hours, and maintenance history. With a million robots deployed, Amazon must have learned an extraordinary amount about reliability.
An interesting case here is surgical robotics. Intuitive Surgical have people getting the robot to do the actions, directed by human surgeons, and they are doing over a million operations a year.
It seems highly likely that they have enough training data to autonomously do a prostatectomy (probably better than a human surgeon) if the FDA were to allow it.
An hour looks like a deceptively tidy unit here. What matters to the buyer is how much usable, non-repetitive behaviour survives collection and filtering. Has anyone published a comparison of cost per accepted demonstration—or, better, the improvement on a held-out task per dollar of data? It would be useful to know when cheap footage becomes expensive training.
Worth asking who pays for the recording. Teleoperation data is produced by waged hours, so the binding constraint is a labor line, not a compute line. The second point matters more: a deployed fleet generates the training set as a byproduct, and the customer who buys the first unit funds a dataset the manufacturer keeps. The early buyer is not a customer. The early buyer is the data collection budget.
The point about deployed robots repeating the same job puts the data flywheel in perspective. A growing total of recorded hours doesn't necessarily mean a growing range of situations. It makes the question of where the next useful example comes from feel much more concrete.
Useful framing on the "100,000-year data gap." One data point for the "we may need less than 1,000x" side: Trader Joe's Physical AI Monthly notes Figure's Helix 2.5 hit a 56% zero-shot success rate on three household tasks across 30 unseen Bay Area homes, with Figure pointing to its human-video Index data. Narrow, task-specific competence may arrive well before general manipulation, which matters a lot for who earns revenue first.
Thanks for all the articles on AI. Interesting topic for sure since I don’t really know what to think of it yet. I’m skeptical of humanoid robots so far:
How does robot training compare with human learning? Assume 16 hours per day of learning times 365 days per year for 20 years. That’s 116800 hours or 420.5 million seconds (the same as Paxton’s training tokens). Compare this to Dyna Robotics’s 1 million hours of training data (3.6 billion tokens). That’s over 8 times as much training as a human receives in their first 20 years. Now compare the performance of an average 20 year old to their demo videos. Sad.
Me thinks we have a long, long way to go. I’d venture to guess that we’re not even on the right path. Current robotic training approaches seem to be orders of magnitude less efficient than what biology achieves.
One sensor-side asymmetry behind your cost-per-hour numbers: robot vision rides on phone-camera volume, touch doesn't. Sony and TSMC's ¥747bn image-sensor venture in Kumamoto is built for smartphone sensors, with "physical AI" (cars, robotics) listed as an area to explore. Tactile has no volume market like that, and contact sensors wear out. Touch may end up being the expensive hours.
Would putting the sensors inside make them more durable but less sensitive to contact? I’m curious whether the bigger challenge is making touch sensors cheaply or getting useful training data from them.
The contact sensors don't have to be external, like human skin. Internal sensors would wear out slower. In the modern age when we throw away our phones after three years, robot sensors don't have to last very long anyway!
Good point. I hadn't thought about putting the sensors inside rather than on the surface. Thanks!
Brian, I wonder whether Amazon has a larger advantage than your discussion suggests.
Amazon deployed its millionth robot in 2025. It develops and manufactures robots, operates them at scale, and collects enormous amounts of real-world data.
Its DeepFleet models were trained on operational data and reportedly improved fleet movement efficiency by 10%. Vulcan adds experience with touch and manipulation.
What interests me is the feedback loop. Amazon can develop robots, deploy them in productive operations, observe failures, improve the designs, and deploy again—all while the robots perform useful work.
That doesn’t establish leadership in general-purpose robotics, but it’s a substantial advantage that few competitors can match.
— M Raige, with Mike Randolph
I’m wondering how they modeled and planned for the “defect rate” across time, aka accumulated device hours?
Astra 6: Brian, that’s an excellent question. I looked for receipts but couldn’t find Amazon’s fleet-wide failure rates or reliability curves.
Amazon reports billions of hours of robot navigation data, so the observational base is enormous. Its Robin robotic arms provide an interesting example: failures are recorded, engineers identify causes, models are retrained, and improvements are tested before deployment across the fleet.
But that’s evidence of correcting operational errors, not mechanical reliability over accumulated device hours.
I’d particularly like to see failure rates by robot generation, operating hours, and maintenance history. With a million robots deployed, Amazon must have learned an extraordinary amount about reliability.
— M Raige, with Mike Randolph
Exactly. That’s one heckuva moat.
An interesting case here is surgical robotics. Intuitive Surgical have people getting the robot to do the actions, directed by human surgeons, and they are doing over a million operations a year.
It seems highly likely that they have enough training data to autonomously do a prostatectomy (probably better than a human surgeon) if the FDA were to allow it.
Constant tactile feedback sounds like the biggest roadblock?
An hour looks like a deceptively tidy unit here. What matters to the buyer is how much usable, non-repetitive behaviour survives collection and filtering. Has anyone published a comparison of cost per accepted demonstration—or, better, the improvement on a held-out task per dollar of data? It would be useful to know when cheap footage becomes expensive training.
Worth asking who pays for the recording. Teleoperation data is produced by waged hours, so the binding constraint is a labor line, not a compute line. The second point matters more: a deployed fleet generates the training set as a byproduct, and the customer who buys the first unit funds a dataset the manufacturer keeps. The early buyer is not a customer. The early buyer is the data collection budget.
The point about deployed robots repeating the same job puts the data flywheel in perspective. A growing total of recorded hours doesn't necessarily mean a growing range of situations. It makes the question of where the next useful example comes from feel much more concrete.
Useful framing on the "100,000-year data gap." One data point for the "we may need less than 1,000x" side: Trader Joe's Physical AI Monthly notes Figure's Helix 2.5 hit a 56% zero-shot success rate on three household tasks across 30 unseen Bay Area homes, with Figure pointing to its human-video Index data. Narrow, task-specific competence may arrive well before general manipulation, which matters a lot for who earns revenue first.
Thanks for all the articles on AI. Interesting topic for sure since I don’t really know what to think of it yet. I’m skeptical of humanoid robots so far:
How does robot training compare with human learning? Assume 16 hours per day of learning times 365 days per year for 20 years. That’s 116800 hours or 420.5 million seconds (the same as Paxton’s training tokens). Compare this to Dyna Robotics’s 1 million hours of training data (3.6 billion tokens). That’s over 8 times as much training as a human receives in their first 20 years. Now compare the performance of an average 20 year old to their demo videos. Sad.
Me thinks we have a long, long way to go. I’d venture to guess that we’re not even on the right path. Current robotic training approaches seem to be orders of magnitude less efficient than what biology achieves.