Language models received an absurd inheritance: humanity had already spent centuries producing text.
Books, websites, manuals, code, forums, academic papers, transcripts, subtitles, documentation, and every other form of written debris existed before modern foundation models arrived. The training corpus was messy, legally complicated, redundant, biased, and enormous.
Robots did not get that inheritance.
There is no equivalent internet containing billions of synchronized examples of a specific robot hand gripping a wet glass, recovering from a slip, noticing that the glass is heavier than expected, adjusting force, avoiding the plate beside it, and setting the glass down without teaching the kitchen a lesson about gravity.
The physical world does not come pre-labeled.
A robot learning from real interaction needs more than images. It may need camera streams, joint positions, motor currents, force and torque data, gripper state, tactile information, language instructions, timestamps, calibration, success labels, failure states, and environmental context.
Then the data must be aligned. If the camera timestamp drifts from the joint-state timestamp, the model learns a slightly wrong relationship between seeing and acting. If a teleoperator compensates for a sticky joint, the dataset may quietly encode a hardware defect as policy.
Every physical dataset comes with a machine attached to it. Machines age.
Open X-Embodiment proved diversity helps.
Google DeepMind and dozens of academic partners assembled the Open X-Embodiment dataset from 22 robot types across more than one million episodes. Their RT-X work showed that models trained across diverse embodiments could outperform narrower models on average and generalize better across tasks.
The important idea is not the exact benchmark score. It is the direction: robot experience can transfer.
That suggests a robotics analogue to foundation models, where many tasks and machines contribute to shared representations rather than every new robot beginning as an amnesiac with expensive elbows.
But embodiment is not just another input format.
A text model can often consume a sentence regardless of which keyboard produced it. Robot data is tied to morphology.
A seven-degree-of-freedom arm and a humanoid with two hands do not share the same action space. One gripper closes with parallel jaws. Another has fingers. Camera positions differ. Joint limits differ. Payloads differ. Control frequencies differ. Coordinate frames differ.
To train across machines, researchers need representations that translate between embodiments: goals, end-effector trajectories, object relationships, contact states, semantic actions, or other abstractions that survive hardware changes.
This is why “just collect more data” is both correct and insufficient.
Real-world data is expensive because failure costs something.
Text generation can run thousands of examples in parallel on servers. Physical training consumes floor space, technicians, spare parts, battery cycles, calibration time, and damaged hardware.
A robot failing to write a sentence produces bad text. A robot failing to place a part can damage the part, the fixture, itself, or the person standing nearby. Safety constraints therefore restrict exploration exactly where learning systems often benefit from exploration.
Simulation helps, but simulation has a tax: the gap between modeled physics and reality. Friction, deformable objects, cable behavior, glare, dust, wear, people moving unpredictably, and cheap manufacturing tolerances all conspire to remind the model which universe it actually lives in.
The data factory may become the strategic asset.
Reuters reported that ACE Robotics is attempting to scale real-world data collection dramatically, including using lightweight sensing systems on production lines. The company's leadership argues that current industry datasets are far below the scale likely required for general embodied intelligence.
Whether ACE's specific timeline proves correct is less important than the industrial logic.
If general-purpose robotics improves strongly with diverse experience, then warehouses, factories, stores, hospitals, and homes become data-generation environments. A company with thousands of deployed robots can collect rare failures and edge cases competitors never see.
That creates a feedback loop:
more deployment → more experience → better models → better reliability → more deployment
The moat is not merely the robot. It is the accumulated interaction history.
This also creates an interoperability war.
If every manufacturer stores actions, sensor states, and task labels differently, the industry fragments its own training corpus. Standardized representations could accelerate everyone, but companies also have incentives to keep high-value interaction data proprietary.
That tension will shape robotics in the same way data formats and APIs shaped computing. Open datasets can lift the baseline. Proprietary deployment fleets can preserve an edge.
Watch the boring tasks.
Humanoid sprint records are excellent demonstrations of actuators, balance, and control. They are also easy for cameras to love.
The commercially decisive tasks are more humiliating: plug in a cable, open packaging that deforms unpredictably, handle transparent objects, separate stuck items, recover when a drawer jams, notice a tool is missing, ask a human for clarification, and resume after interruption.
Those are not one-shot tricks. They require perception, memory, uncertainty, planning, dexterity, and recovery.
A robot that can perform a dazzling backflip in a prepared space and fails three times while loading a dishwasher is not contradictory. It is a reminder that dynamic control and general manipulation are different capability stacks.
What this analysis cannot establish yet.
No public evidence establishes a universal data threshold at which general-purpose robotics suddenly emerges, and the analogy to language-model scaling may break in important ways. Better simulation, model architecture, active learning, human demonstration, hardware standardization, or task-specific priors could reduce the amount of real-world data required. The narrower assessment is that data collection, normalization, and deployment feedback deserve to be treated as first-order engineering infrastructure rather than as an afterthought.
The embodied-AI race is likely to be constrained as much by data infrastructure as by motors or model architecture. Physical data is expensive, embodiment-specific, and difficult to standardize. The organizations that learn to collect, normalize, share, simulate, and continuously harvest real-world experience may compound faster than those optimizing demonstrations alone.
The metric that matters
Stop counting promotional videos. Track autonomous task-hours between human interventions, recovery from unexpected states, transfer to unseen tasks, and the diversity of environments represented in training.
When those curves start moving sharply, the “ChatGPT moment” in robotics will not need a keynote to announce itself. The machines will simply stop requiring us to prepare the room for them.