Humanoid Robots: Training Data Costs Just Fell 200-Fold

The consensus has the wrong frame

The bottleneck in humanoid robotics has changed address: today it is called cost per training trajectory.

The consensus is still watching hardware, actuator prices and the GPU-hours behind the foundation model. That is the wrong figure. The figure that predicts parity with manufacturing labour is the cost of robot training data per single executable example, touch included.

Anyone treating the scarcity of robot data as permanent is pricing in a bottleneck that is on the move. We have seen this before with training silicon and with packaging: every scarcity lasts one cycle. Whoever writes it into contracts as permanent pays the full price of the next cycle.

The curve: three points, one order of magnitude per jump

The trajectory of robot data has three documented stages, and each stage shifts the order of magnitude.

• 2022, the teleoperation era: Google's RT-1 dataset required 13 robots and 17 months of collection for roughly 130,000 episodes, according to the public figures in the original paper.
• 2023, the aggregation era: Open X-Embodiment pooled more than one million episodes from 22 robot types, combining datasets from dozens of labs.
• 2026, the human video era: 50 video demonstrations generate 10,000 simulated trajectories, a 200-to-1 ratio, according to UniDex-ViTac^[1].

The first jump multiplied volume by aggregating what already existed. The second jump is different in kind: it changes the source of the data. A teleoperator produces one episode every few minutes, a person in front of a camera produces a video in a few seconds, and simulation does the rest.

The cost curve says this: the human cost per trajectory drops by two orders of magnitude in a single technological step. The 200-to-1 ratio is the number to frame on the wall.

It is the same profile as solar and as inference: the technology matures for years, then a change of source compresses the cost in one stroke.

What UniDex-ViTac demonstrates

The UniDex-ViTac^[1] paper, published on arXiv on 15 September 2026, starts from 50 human videos across ten objects and derives 10,000 simulated trajectories to train a single generalist policy.

The mechanism matters more than the number. Residual reinforcement learning specialists adapt each annotated human interaction to a robotic arm-hand system. Every successful rollout carries the fingertip contact observations with it, and touch, the most expensive data to collect on real hardware, is born in simulation at marginal cost.

The physical results are the part a CTO needs to read twice. The policy, with no real-robot demonstrations and no fine-tuning, succeeds in 73 of 110 trials (66.4%) across six seen objects and five never seen before. The baseline relying solely on the point cloud stops at 60 of 110 (54.5%): a gap of 11.8 percentage points, as the paper^[1] reports. In simulation the gap is similar, 68.3% against 55.5%.

Does 66% clear the industrial threshold? Not yet. The point is a different one: that 66% was bought with 50 videos, and the marginal cost of the next 66% is close to zero.

Why the bottleneck migrates

Every scarcity lasts one cycle.

Training silicon looked like the limit in 2023. Then the constraint moved to packaging, with CoWoS capacity rising from 13,000 to 120,000 wafers a month as this desk has already measured, then to HBM memory. Robotics follows the same pattern: actuators, then on-board compute, then data.

The data constraint has one extra property: it is the most elastic of the three. An actuator requires a factory, a simulated trajectory requires a GPU-hour, and the price of a GPU-hour has been falling for years. Once the source becomes human video, supply is unlimited by definition: billions of hours of manipulation already on film.

On-board compute is running in the same direction. Arm is bringing the topic of scaling physical AI to the RoboBusiness stage, as The Robot Report^[2] previews. When the supplier of device architectures talks about scale, local inference is already a volume problem and has stopped being a feasibility problem.

Cliff event: cost parity in 2028-2029

Cliff event: tactile trajectories generated from video at marginal cost, in 2027, with manipulation policies trained for a specific production line in days rather than months.

This desk's position remains as stated: cost parity between a humanoid and a manufacturing worker in 2028-2029. The first anchor is BMW at 25 dollars per robot-hour, and the curve tracks solar, down 90% in ten years.

The consensus says 2035. The difference is six years, and the line item that closes the gap is data itself. Hardware and compute were already on known curves. Data was the only variable priced as flat, and the 200-to-1 ratio makes it steep.

This goes beyond a trend: it is a regime change. Inevitable for years. From this paper onwards, imminent too.

Three categories that will change shape by 2028

Three categories that, by 2028, will exist in a form different from today's:

1. Teleoperation fleets for data collection. Their asset is the human cost per episode: when that cost falls 200-fold, the business model migrates towards physical validation of policies and abandons collection.
2. Robot dataset licences. A dataset of teleoperated episodes is worth what an oil field is worth: a great deal, until someone finds a free source. Human video is that source.
3. Multi-year contracts with humanoid OEMs that sell proprietary data as a competitive advantage. The moat shifts to the video-simulation-policy pipeline and to the on-board tactile sensor.

For the CTO: reassess the simulation stack and the contact sensors now, before it becomes obvious. For the VC: the bet that looks impossible is the team selling video-to-policy conversion, never the team selling teleoperation hours. For procurement: every three-year contract indexed to a flat cost of robot data needs renegotiating with step-down clauses.

For the Chief Strategy Officer, the risk is a three-year plan that assumes a scarcity already dissolving.

Capital locked into data-collection fleets behaves like capital locked into CoWoS capacity in 2024: it earns as long as the curve allows, then it becomes sunk cost.

What would change my mind

The thesis has a weak point and I will name it: the sim-to-real gap.

UniDex-ViTac's 66.4% physical result comes on eleven tabletop objects. An assembly line has tolerances, flexible cables and insertion forces that simulation reproduces less well. Should video-born policies remain below teleoperated ones on real industrial tasks for another two years, cost per trajectory would stop being the right metric. My curve would lose its third point.

The second doubt is touch. Four binary contact signals are enough to grasp. Screwing and cabling require continuous forces, and there the gap between simulated sensor and real sensor is wider.

Ninety percent of analysts are right about the present. They are wrong about the pace of change, and the pace, here, is a 200-to-1 ratio already measured and already published.

Forecast, horizon, kill signal

Forecast: by 31 December 2027, at least one of Figure, 1X, Apptronik and Tesla Optimus will publicly disclose the new source of its data. The statement will say that more than 50% of the training data for its manipulation policies comes from human video or simulation, rather than from direct teleoperation.

Confidence: Medium, 65 out of 100. Horizon: 471 days. This is a forecast about market timing, which is why it stays at medium confidence. The forecast about the technology, the collapse of cost per trajectory, is high confidence.

Kill signal: on 31 December 2027, every public disclosure from the four manufacturers still names human teleoperation as the majority source of manipulation training data. In that case the data bottleneck was harder than the curve suggested, and cost parity slips beyond 2029.