Humanoids and General-Purpose Robots

For decades, humanoid robots were the realm of science fiction and expensive research demonstrations. Honda's ASIMO could walk up stairs and wave at press conferences, but it took minutes to plan each step and fell over at the slightest disturbance. The gap between demonstration and deployment seemed vast. Then, between 2021 and 2024, something shifted. A convergence of better actuators, improved perception, more capable AI, and enormous private investment produced a new generation of humanoid robots that could fold laundry, plug in cables, carry boxes, and navigate unstructured spaces. The race to build the first genuinely useful general-purpose humanoid robot became one of the most consequential technological competitions of the decade.

Why Humanoid Shape?

The argument for a humanoid form factor is not sentimental — it is pragmatic. The world is built for human bodies. Doorknobs, ladders, tool handles, vehicle controls, kitchen appliances, hospital equipment, construction scaffolding, and office furniture are all designed around the dimensions, reach, grip, and locomotion of a human body. A robot that matches that morphology can, in principle, operate in any environment built for a person without requiring any modification to the infrastructure. This stands in contrast to special-purpose robots, which are designed to excel at one task in one environment. An automotive welding robot is extraordinarily fast and precise, but it cannot carry a box to a different room. A warehouse picking robot can locate and retrieve a specific item from a shelf, but it cannot climb a ladder to stock a high shelf or navigate a crowded sidewalk. The promise of the humanoid is generality: one robot, any task, any environment built for people. There is also a training-data argument. Decades of research in computer vision, natural-language processing, and embodied AI have produced datasets recorded by cameras placed at human eye level, from the point of view of a creature with two hands and two legs. A robot that matches this morphology can potentially leverage all of that human-generated data for its own learning — a humanoid robot watching a cooking video is watching something recorded by a body like its own.

Key Players and Systems (2022-2025)

Boston Dynamics' Atlas has been the most publicly visible humanoid platform for over a decade. The current all-electric version can run, jump, backflip, and manipulate objects with impressive dexterity. Boston Dynamics uses Atlas primarily as a research and demonstration platform, though the company announced plans to commercialize it for industrial tasks. Tesla's Optimus robot was unveiled in 2022 and has iterated rapidly. By 2024, Optimus Gen 2 demonstrated folding laundry, sorting objects by color, and performing assembly tasks in Tesla's own factories. Elon Musk's stated goal is to produce millions of Optimus units, enough to have one in every household. The robot uses Tesla's FSD (Full Self-Driving) hardware and neural processing stack. Figure AI's Figure 01, in partnership with OpenAI, demonstrated in March 2024 a robot capable of natural conversation and real-time task completion based on spoken instructions. The demo showed Figure 01 locating an apple, picking it up, and handing it to a human who asked for something to eat — reasoning through the request in natural language and executing it physically. Agility Robotics' Digit is designed specifically for warehouse and logistics environments. Amazon announced a pilot deployment of Digit in its fulfillment centers. Digit is notably not a full humanoid — it has no arms designed for fine manipulation — but it can walk on two legs and carry totes. Unitree Robotics, a Chinese company, produces the G1 and H1 humanoids at significantly lower price points than Western competitors, raising questions about the competitive landscape of the market. The common thread across all these systems is the use of machine learning — especially imitation learning and reinforcement learning — to replace the painstaking hand-coded controllers of the previous generation.

The Hard Problems: Dexterity, Navigation, and Generalization

Despite impressive demonstrations, humanoid robots in 2025 face three hard problems that no team has fully solved. Dexterous manipulation is the first. The human hand has 27 degrees of freedom, tens of thousands of mechanoreceptors in the fingertips, and a control system developed over millions of years of evolution and twenty years of individual learning. Current robotic hands have far fewer sensors and actuators, much lower control bandwidth, and nowhere near the grasping versatility of human hands. Threading a needle, tying a surgical knot, or feeling whether a fruit is ripe are trivially easy for a human and extremely hard for a robot. Robust locomotion in unstructured environments is the second. Bipedal walking is statically unstable — a standing human is always falling forward and catching themselves. Dynamic balance across stairs, rubble, ice, and push disturbances remains an unsolved engineering problem at the reliability level required for deployment in safety-critical environments. Generalization across task domains is the third. Current demonstrations show robots performing specific tasks learned from hundreds or thousands of demonstrations. The goal of a truly general robot — one that can be given any reasonable household or workplace task and figure out how to do it — remains distant. The combinatorial space of tasks, objects, environments, and instructions is far larger than any training dataset that has been assembled.