In 2012, when Gill Pratt launched the DARPA Robotics Challenge (DRC), the ambition was explicit: replicate the catalytic effect that earlier DARPA Grand Challenges had produced for autonomous vehicles, but this time for humanoid robots. The DRC offered teams a series of disaster-response tasks — driving a vehicle, opening doors, climbing stairs, cutting through walls — designed to push bipedal machines beyond the confines of laboratory demonstrations. The results were a mix of genuine engineering milestones and a now-legendary reel of robots toppling mid-task, a spectacle that became one of the most-watched robotics videos on the internet. The contrast was instructive. Boston Dynamics' Atlas emerged as a credible platform. But the broader field revealed how far humanoids remained from operating in unstructured environments without human babysitting.
A decade after the DRC concluded, the humanoid sector is experiencing a second surge of attention and capital. Pratt, now leading the Toyota Research Institute, brings to this moment the perspective of someone who watched the first wave crest and recede. His assessment is notably sober: while advances in machine learning and actuator design have shifted the landscape, the core engineering problems — energy efficiency, robust balance, and the ability to generalize learned behaviors across novel situations — have not been solved so much as reframed.
From blooper reels to balance sheets
The original DRC was, in many respects, a stress test for an entire class of machines. The tasks were deliberately chosen to simulate real disaster scenarios, environments where wheeled robots would struggle and human rescuers would face unacceptable risk. The competition demonstrated that humanoid form factors could, in principle, navigate spaces built for people. It also demonstrated that the gap between "in principle" and "reliably" was vast.
What has changed since then is not a single breakthrough but an accumulation of incremental gains. Modern actuators are lighter and more energy-dense. Simulation environments allow teams to train control policies on millions of virtual falls before a physical prototype ever touches the ground. And the integration of large-scale machine learning into perception and planning pipelines has given robots a degree of environmental awareness that was unavailable to DRC competitors. These improvements are real, but they compound rather than eliminate the underlying difficulty of making a tall, narrow-based machine move safely through a world full of uneven surfaces, unexpected obstacles, and fragile objects.
The commercial question is whether these compounding gains have crossed a threshold. Several well-funded startups and established players are now positioning humanoids for logistics, manufacturing, and warehouse operations — domains where the environment can be partially controlled and the economic case for automation is already established. The humanoid form factor carries a specific advantage in these settings: it can, in theory, use the same tools, walkways, and workstations designed for human workers, avoiding the costly infrastructure retrofits that specialized automation demands.
The intelligence problem, not the hardware problem
Pratt's framing suggests that the decisive bottleneck is shifting from hardware to intelligence. Building a humanoid that can walk is no longer the primary challenge; building one that can decide what to do next in a cluttered, changing environment is. This distinction matters because it redefines the competitive landscape. The companies most likely to field commercially viable humanoids may not be those with the most elegant mechanical designs, but those with the most effective learning architectures and the richest training data.
This echoes a pattern familiar from the autonomous vehicle industry, the very domain DARPA's earlier challenges helped create. In self-driving cars, the hardware problem — sensors, compute, vehicle platforms — was largely addressed within a few years of serious investment. The intelligence problem — handling edge cases, rare scenarios, and the long tail of real-world driving — consumed the next decade and remains incompletely solved. Humanoid robotics may be entering an analogous phase, where the visible hardware inspires confidence but the invisible software determines the timeline.
The tension, then, is between the pace of expectation and the pace of engineering. Capital markets and corporate strategy departments are pricing in a future where humanoids perform useful labor at scale. The machines themselves are still learning not to fall down. Pratt's career arc — from launching a competition designed to accelerate the field to leading a research institute focused on the patient, methodical work of making robots genuinely capable — may itself be the most accurate map of the road ahead. Whether the current wave of investment can tolerate that timeline is a question the market has not yet been forced to answer.
With reporting from IEEE Spectrum Robotics.
Source · IEEE Spectrum Robotics
