For decades, the field of robotics suffered from a persistent gap between ambition and utility. While researchers envisioned machines capable of navigating the complexities of human environments — aiding the elderly, performing hazardous tasks, or managing household chores — the reality was largely confined to the repetitive precision of auto-plant assembly lines. The industry aimed for the versatile androids of science fiction and arrived, instead, at the Roomba. This history of over-promising and under-delivering left venture capital wary of the sector for years, channeling funding toward software and digital platforms where returns were faster and more predictable.
That hesitation has evaporated. In 2025, investors poured $6.1 billion into humanoid robotics, a fourfold increase over the previous year. The surge is not driven by a sudden improvement in hardware — actuators, sensors, and batteries have advanced incrementally — but by a fundamental shift in how machines are taught to interact with the physical world. The industry is moving away from the brittle, rule-based programming of the past toward the same large-scale learning models that have revolutionized digital intelligence.
From instruction sets to learned behavior
The traditional approach to robotics treated every task as an engineering problem to be solved exhaustively in advance. Programming a robot to fold a shirt, for instance, required an explicit list of instructions: calculating fabric deformation, identifying collars and seams, and adjusting for every possible rotation and material type. This "if-then" logic could achieve remarkable precision in controlled factory environments, where objects arrive in known orientations on predictable conveyor belts. But it fails the moment it encounters the unpredictability of a real home, a cluttered warehouse, or any setting where conditions vary from one second to the next.
The shift now underway reframes physical movement as a data problem rather than a geometric one. Instead of hand-coding every contingency, researchers train models on large datasets of robotic interaction — demonstrations, simulations, and real-world trial runs — allowing the system to develop generalized motor policies. The approach borrows directly from the paradigm that produced large language models: scale the data, scale the compute, and let statistical patterns do the work that explicit rules cannot. In language, this method turned rigid chatbots into fluid conversational agents. In robotics, the promise is analogous — turning the static tools of the factory into adaptable agents capable of operating in unstructured environments.
The parallel is instructive but imperfect. Language models operate in a domain where errors are low-cost: a poorly constructed sentence carries no physical consequence. A robot that misjudges the force needed to grip a glass or miscalculates a step on a staircase operates under far less forgiving constraints. The gap between digital fluency and physical reliability remains one of the central engineering challenges in the field.
What the capital is chasing
The $6.1 billion figure reflects a bet not just on technology but on timing. Several converging factors have made humanoid robotics investable in a way it was not five years ago. Simulation environments have grown sophisticated enough to generate useful training data at scale, reducing the cost and time of real-world experimentation. Transfer learning techniques allow models trained in one context to adapt to adjacent tasks without starting from scratch. And the labor economics in sectors like logistics, elder care, and light manufacturing have shifted enough to make the unit economics of a general-purpose robot more plausible than before.
Yet the history of robotics counsels caution. The field has seen previous waves of enthusiasm — the Japanese humanoid programs of the early 2000s, the DARPA Robotics Challenge of 2015 — that generated excitement and funding without producing commercially viable products at scale. Each wave advanced the underlying science, but each also revealed how far the gap between laboratory demonstration and reliable deployment truly stretched.
The current moment is different in at least one structural respect: the AI models being applied to robotics are improving on a curve driven by investment across the entire technology sector, not just robotics alone. Roboticists are, in effect, riding a wave of capability funded primarily by the demand for digital AI. Whether that borrowed momentum translates into machines that can reliably operate in the physical world — or whether the old gap between ambition and utility reasserts itself — remains the defining question for the billions now flowing into the sector.
With reporting from MIT Technology Review.
Source · MIT Technology Review
