What happened
In a perspective published in *Science Robotics*, Ken Goldberg asks whether data alone will produce general-purpose robots, as they have helped advance language and other AI fields. He accepts that model-free, end-to-end AI may eventually enable general-purpose robots, but questions how long that will take given the data available today.
Goldberg compares the scale of training data for language and robotics. Using commonly accepted conversions from words and image tokens into time, he estimates that contemporary vision-language models use internet-scale text and image data equivalent to roughly 100,000 years of human reading or viewing. Robot training data, by contrast, pair video inputs with robot motion commands, and Goldberg says those paired data do not exist on the internet. He describes teleoperation—people remotely controlling robots to perform tasks repeatedly—as one collection method. The largest reported dataset he cites is on the order of one year of data, collected in less than a year across many human-robot systems. Extrapolating from the scale difference, he says a ChatGPT-sized set of robot data could take 100,000 years to collect at current rates.
He discusses simulation and internet video as two other possible data sources. Simulation can help with tasks such as flying, walking, and backflips, but he says it remains unreliable for manipulation because real contact involves changing forces, friction, and small deformations. Video sites contain substantial footage, but recovering precise three-dimensional finger and object motion from video is, in his account, a major computer-vision challenge not expected to be solved soon.
Goldberg points instead to commercial robots working with real objects in real environments. He cites Waymo robot taxis, which operate in several US cities and can connect to remote human operators when they need help, and Ambi Robotics’ package-sorting and stacking machines. He says both companies have created data flywheels: operational data improve performance and support adjacent skills. Waymo’s example is highway merging; Ambi’s is package stacking, distinct from sorting.
The article calls the combination of rigorous engineering and AI “good old-fashioned engineering,” or GOFE. It names Kalman filters, Gaussian processes, inverse kinematics, and motion planning as examples. Goldberg says modular, model-based methods can be tested, fixed, and fine-tuned independently, then combined with model-free methods to reach reliability levels needed in commercial environments. He reports that Ambi Robotics collected 22 years of real robot data over four years while sorting more than 100 million packages, using model-based and model-free methods, and is using those data to train later model-free systems. The article does not describe a new Nvidia system or a new product launch.
What the archive adds
New Runtime’s earlier coverage has tracked robotics hardware and training-data developments, including NVIDIA’s GR00T 2 and Isaac Lab 3.0 on 2026-11-09 and a basket of hardware-agent developments on 2026-10-31. Goldberg’s article is not a new event; it adds an earlier argument and an Ambi Robotics data example to that developing thread.
Trend signal
This supports a practical reading of the shift toward embodied training data: operational deployment can generate data that are difficult to obtain through internet scraping or simulation. The article’s Ambi example connects useful production work to later model improvement. This is context for, rather than proof of, a general-purpose robotics path. Agent economics moves to completed work is not a close fit: the article discusses data and reliability, not cost per verified task.
Why now
The article itself frames the timing around whether the public and investors will wait for general-purpose robots. Its answer is to build toward them through useful, bounded deployments that produce real data, rather than wait for a massive robot dataset to appear first.
Why it matters
For teams building robots, the article challenges a simple “scale the model and wait” strategy. If real-world paired sensor and action data are scarce, a deployed system’s ability to work reliably on a bounded task can matter as much as its model architecture. Engineering modules can make behavior testable and fixable; a commercial deployment can then expose failures and generate data in the conditions where the robot must operate.
That is a different path from treating simulation, internet video, or teleoperation as interchangeable sources of training material. Each has a limitation Goldberg identifies: simulation’s mismatch for manipulation, video’s difficulty in recovering precise motion, and the labor involved in teleoperation. The operational flywheel he describes depends on a task with enough repeated real-world activity to generate useful examples, plus methods to convert those examples into measurable performance improvements. It does not establish that every commercial robot will produce transferable data or that narrowly trained systems will become general-purpose. But it gives builders a concrete design question: can the robot’s useful work also create evidence and data that improve the next version?
Limits and uncertainty
Goldberg’s article is a perspective, not a controlled comparison of competing robotics strategies. The “100,000 years” figure is an order-of-magnitude analogy based on stated conversions, not a measured timeline for building a general-purpose robot. The article does not disclose Ambi Robotics’ data quality, collection costs, evaluation protocol, or the gains achieved by training on the reported 22 years of data. It also does not independently establish how much of the claimed flywheel transfers to new tasks or environments. The examples illustrate the argument; they do not prove that this approach will close the gap for robotics as a whole.
Skill to build now
For a robot operating in a repeated commercial task, instrument the workflow so each run records the relevant observations, actions, outcome, and any human intervention. Keep the task boundary explicit, and evaluate whether the recorded examples improve a held-out version of that same task before claiming a broader capability. This makes the proposed data flywheel testable without assuming that more data automatically means more generality.
Reader takeaway
The article’s central lesson is not that engineering replaces learned models. It is that engineering can help robots do useful work before robot-scale datasets exist—and useful work may itself become a source of training data. Treat that as a deployment strategy to test, not as evidence that general-purpose robotics is imminent.