Humanoid robots struggle with basics despite massive investment

A version of this article appeared on Bloomberg News.

The gap between viral marketing and the functional reality of humanoid robots remains a significant hurdle for industries reliant on complex physical labor. While recent demonstrations show machines performing backflips or dancing, these controlled displays often mask a lack of basic reliability required for real-world applications.

Billions of dollars have flowed into the sector, fueled by the promise of "physical AI" that could eventually supplement human labor in factories and on construction sites. However, engineers and researchers suggest that the dexterity required for even simple manual tasks is proving much harder to replicate than the language fluency seen in digital AI models.

Ken Goldberg, a roboticist at UC Berkeley, recently highlighted what he calls a 100,000-year data gap. This refers to the lack of physical interaction data available to train robots compared to the vast repositories of text used to train large language models. Without this data, robots struggle to navigate unpredictable environments.

For sectors like construction, the limitations are particularly acute. While a robot might be programmed to walk across a flat warehouse floor, the uneven terrain and constantly changing variables of a building site present a level of complexity that current humanoid systems cannot yet manage safely or efficiently.

The challenge is not just movement, but manipulation. Tasks performed by a plumber or an electrician require fine motor skills and the ability to adapt to unique physical constraints. Research published in Science Robotics suggests that these blue-collar trades remain some of the safest from automation for the foreseeable future.

Investors are increasingly concerned about a potential bubble. Data from CB Insights indicates that while industrial and logistics robots are already generating revenue in predictable environments, humanoids have yet to prove their commercial value. Many startups are now under pressure to show results beyond lab-based prototypes.

In China, the government has already begun advising the industry to balance development speed against the risk of speculative bubbles. The concern is that the hype surrounding humanoid capabilities is far outstripping the actual physical hardware and software integration currently available to the market.

Despite these warnings, firms like Boston Dynamics and various Silicon Valley startups continue to push for breakthroughs in reinforcement learning. These efforts aim to shorten the time it takes for a machine to learn a new task, though most experts agree that widespread deployment in homes or complex workplaces remains years away.

For now, the most practical applications remain confined to highly structured environments. Warehouses and simple factory assembly lines offer the predictability required for current systems to operate. Until robots can master the basics of human-like dexterity, their role in more dynamic industries will remain limited to experimental trials.