Robotics’ data drought still stalls progress

By late 2026, robotics experts were treating data scarcity as a bottleneck that earlier scaling efforts had not solved. By 2022, UC Berkeley’s Bridge Data and Google’s RT-1 had already shown scalable teleoperation with consumer VR hardware and fleets of robots, collecting tens of thousands to over 130,000 demonstrations and shifting imitation learning from bespoke collection to centralized datasets, but that progress still did not close the gap between fast-moving AI methods and dependable embodied performance.

The late-2026 diagnosis was reinforced by multiple voices and research efforts. In Cognitive Revolution’s September discussion, the speaker said the “big breakthrough came in 2023” when researchers could train “a language a VLM” to output robot actions directly, and noted that “most…modern robotic stacks…are based on that same basic paradigm,” while still pointing to limits in context and generalization. The same talk linked the RT2/VLA lineage to startups, including physical intelligence and Generalist, and AI Engineer argued that teleoperation still follows “the exact same principles” as a 1957 demo, underscoring the widening gap between 2 model capabilities and physical robot hardware reliability.

By late 2026, the diagnosis was that robotics cannot scale on the same recipe that worked for language. On Lenny’s Podcast, Fei-Fei Li said robot training lacks actions in 3D worlds, so robotics needs brains, physical bodies, and application scenarios; she pointed to self-driving’s 20 years from prototype to Waymo as evidence that real-world iteration is slow. On We Study Billionaires, the missing manipulation pipeline led to the blunt claim that if you add up all those tokens, you end up with 100,000 years, and that robotics is 100,000 years behind the language model.

What sharpened the consensus was that neither more robots nor better models seemed enough without better embodied coverage and sturdier execution. TheAIGRID described Yann LeCun’s point that humans are preloaded with 17 years of embodied experience, while Sergey Levine warned on Invest Like the Best that putting 20 robots in a room to learn together still struggles with tail cases and edge cases; Y Combinator added that no two robot platforms are the same and platforms drift, limiting reuse. Even newer systems only reached 74.2% and 82.2% benchmark success, and Rod Brooks, after 40, 60, and 50 years in the field, said the problem is much more nuanced than it seems.

Teleoperation still looks less like a scalable data engine than a labor process. On Gradient Flow, the pipeline was defined bluntly: “there's a robot and a human will control robot to do specific task,” with the “10 to 20 hour of data” “generated by human doing the task piloting task but on a robot”; that same discussion contrasted a “pre-training stage that's training on webcale data” with teleoperation, where “the 10 to 20 hours is post training,” underscoring how little embodied experience this human-driven stage contributes relative to what modern learning systems usually consume.

The bottleneck is not only that every example needs a person, but that the workflow itself sheds throughput. Suen said on AI Engineer that teams first lost “two or three minutes for each episode to be encoded,” and after changing the process still “had to wait for good 30 40 minutes before we start the next session”; meanwhile, practical pipeline comparisons showed “3,200 HiFi-UMI trajectories per task set against approximately 300 teleoperation trajectories,” and another AI Engineer explainer noted that even “about 10,000 hours of teleyop” still means “a human is moving the robot around” to make training data.

Robot bodies remain the rate limiter because physical execution has not kept pace with gains in reasoning. As Chelsea Finn told Fast Company, even when a robot identifies the right subtask, “translating that into low-level motor commands is incredibly difficult,” and in Possible she described “months of 0% success rates on laundry folding before a single architectural insight unlocked the capability,” because “there are so many different ways for even just a single shirt” to be crumpled, exposing how brittle dexterous control still is once models leave tidy demos and enter contact-rich reality.

The hardware gap is most obvious in endurance and hands: Prof G Media said “we are not currently able to build a humanoid robot that can run for more than four hours,” with “Most humanoids can support only two to four hours of charge — not enough to make it through an eight-hour shift,” while the same analysis called out the “hand problem,” noting human hands’ “27 individual bones, dozens of tendons, and over 17,000 nerve endings” are hard to replicate. That helps explain why Rodney Brooks wrote that “We are more than ten years away from the first profitable deployment of humanoid robots even with minimal dexterity.”

Benchmarks on real manipulation show why long-horizon deployment still stalls at the body. The Good Tech Companies wrote that “Dex Terra T… coordinating contact rich manipulation across them is still unsolved at production reliability,” with performance dropping from around 80% “toward zero on contact rich tasks like standing a cup upright,” and that “The best policies on real multi finger hands land near 63% average success across a handful of everyday tasks,” reinforcing the AI Engineer view that “everything around robotics has progressed but robotics is still stuck in the same land from the last 70 years.”

The scale gap shows up even in the field’s biggest bets: The Robot Report’s headline, “Rhoda AI exits stealth with $450M to train robots from video,” is less a victory lap than a measure of how costly robotics data remains. If training from video still demands a $450M capital injection, robotics is plainly not operating in the low-friction regime that let language and vision feast on abundant, already-accessible corpora; as one Robotiq speaker put it, “we were gathering data for 40 years in the internet… It is out there,” a contrast that physical AI still cannot claim.

Even the strongest new pretraining efforts remain small and bespoke by comparison: Xiaomi-Robotics-1 reports “over 100k hours of real-world manipulation trajectories collected via UMI devices,” plus a “scalable auto-labeling pipeline” that annotates clips with natural language, and says it “consistently improves with increased data scales and model sizes during pre-training.” That is progress, but it also shows how much extra machinery robotics needs before data becomes usable, and why scale is still constrained: “We Open source the Pi05 model, and people are actually using it and building on it… But at the same time… With robots, we're not there yet because you need a robot in order to interact with it.”