A humanoid robot is a robot whose physical form approximates the human body, typically including a head, torso, two arms with hand-like end effectors, and two legs, a design intended to let the machine operate in spaces, tools, and workflows built for people without requiring the environment to be redesigned. The category spans decades of research prototypes, including Honda's ASIMO, unveiled in 2000, and Boston Dynamics's Atlas, first developed for DARPA in 2013, but attracted a new wave of commercial investment starting around 2022, driven by advances in Reinforcement learning-based locomotion, cheaper actuators and batteries, and the emergence of general-purpose Vision-language model systems capable of interpreting instructions and scenes.
Historical development
Early humanoid research prioritized locomotion and balance as the central hard problem: ASIMO demonstrated stable bipedal walking and stair-climbing, while Atlas, initially built for disaster-response tasks under a DARPA program, became known for dynamic, athletic movement such as running and backflips developed by Boston Dynamics. These platforms relied heavily on classical control theory and model-based planning rather than learned perception, and were largely research or demonstration systems rather than commercial products, reflecting the broader gap between AI's digital and physical progress discussed in the article on Robotics.
The 2022-2026 wave
Commercial interest in humanoids accelerated sharply from 2022 onward. Elon Musk introduced Tesla's Optimus project in 2022, positioning it as a mass-manufacturable general-purpose worker robot and a long-term extension of Tesla's Self-driving car perception stack. Figure AI, backed by investors including OpenAI, Microsoft (AI), and NVIDIA, began piloting its humanoid in warehouse and manufacturing settings starting in 2023 and struck a widely publicized collaboration with OpenAI to integrate large-model reasoning into its robot. Other entrants included Agility Robotics's Digit, already deployed in limited logistics trials, and a wave of Chinese manufacturers such as Unitree and Fourier Intelligence, which drove down hardware costs. Boston Dynamics unveiled a fully electric version of Atlas in 2024, replacing its hydraulic predecessor and shifting its research emphasis toward learned, rather than purely engineered, control.
Vision-language-action models
The central technical bet behind the 2024-2026 wave is that the same recipe that produced capable Large language model systems, pretraining a large model on broad data and then adapting it to a task, can be extended to physical control through vision-language-action, or VLA, models: a Vision-language model backbone is trained or fine-tuned to output robot joint commands directly, allowing a single model to follow natural-language instructions across a range of manipulation and locomotion tasks after training on demonstration data collected via teleoperation. Google DeepMind's RT-2 and successor efforts, along with similar work from Figure, Physical Intelligence, and other robotics-focused labs, are examples of this approach, which its proponents argue could let humanoid robots generalize the way large language models generalize across text tasks, though the amount of real-world physical training data available remains far smaller than the text corpora used for language models.
Outlook and skepticism
Proponents frame humanoid robots as a potential trillion-dollar labor-market shift, capable of performing manufacturing, logistics, and eventually domestic tasks at scale. Skeptics note a long history of overpromised robotics timelines, point to unresolved challenges in dexterous manipulation, battery life, and cost, and argue that task-specific robot designs will often outperform general humanoid form factors for a given job, making the pace and eventual scale of humanoid deployment through the late 2020s genuinely uncertain.