Robots take many forms. Drones have proved remarkably popular for various kinds of surveillance and inspection. Robotic arms are ubiquitous in advanced manufacturing. Warehouse robots are also increasingly popular—some resembling automated pallets. Experimental robots include those that use soft materials and tiny flying robots that operate in swarms.

Ask a child to draw a robot, however, and they will likely draw a humanoid. This fascination has a long history—perhaps dating back to the automatons of the 18th and 19th centuries—and countless science fiction stories focus on humanoid robots. Interest in humanoid robots is also practical. Although non-humanoid robots have some advantages (humans can't fly or climb inside aircraft engines), most built environments are designed around the human form. Humanoid robots, in other words, could plausibly perform "general-purpose" applications.

Although robot researchers have worked on humanoid robots for decades, commercial products have been rare and, where they have existed, largely served only as tools for other researchers. Notably, Boston Dynamics says it has no intention of commercializing Atlus—arguably the most advanced humanoid robot in existence.

Now, some signs suggest that humanoid robots could soon see more meaningful commercialization. Several firms even hope to deliver humanoid robots to the mass market. Examples of companies making technical and commercial progress with humanoid robots follow.

• Figure: Figure is a Silicon Valley start-up targeting general-purpose humanoid robotics. The firm has over $100 million in funding and its engineering team includes former employees of robotics teams at Google, Tesla, and Boston Dynamics. Figure recently demonstrated a robot that can walk and has a five-digit, human-style hand. Although the hand is not fully functional and the robot still walks with bent knees, the team achieved the prototype in a year—much faster than, for example, development of Atlus at Boston Dynamics.
• Agility Robotics: Agility Robotics has begun selling Digit—a humanoid robot that can walk and manipulate basic objects (for example, lifting boxes). The firm is targeting logistics applications and is beginning to find commercial success. In particular, both Amazon and Ford are testing Digit for warehousing applications. Perhaps most significantly, Agility is preparing for the mass market. Agility Robotics recently completed a 70,000‑square-foot factory in Oregon that it says will be capable of producing 10,000 Digit robots a year (though its initial production runs will be in the hundreds).
• Fourier Intelligence: Fourier Intelligence in Singapore recently released a new video showing its production facilities for its GR‑1 humanoid robot. The company says it plans to ship 100 robots to customers by the end of 2023. Right now, those "customers" are R&D partners but GR‑1 appears to have significant potential. Notably, the robot can carry objects of up to 50 kilograms (110 pounds).
• Tesla: Tesla continues to make progress with its Optimus humanoid robot—though it remains an in-house research project rather than a product. Optimus uses the same AI system that Tesla cars use to turn video input into control output. In its latest video of Optimus, Tesla shows the robot accurately locating its limbs, performing yoga poses, and sorting colored blocks.

Key to the potential progress of humanoid robots is advances in AI that could accelerate the ability of such robots to learn. In particular, roboticists are having some success with applying variations of large language models (that power leading generative AI systems like GPT 4) to robotics.

Toyota, in partnership with Columbia Engineering and MIT, recently claimed a "bit of a breakthrough" with a new robotic learning approach called a "Diffusion Policy" that teaches itself to perform tasks by watching humans perform them, in a similar manner to how large-language models learn from human writing. Toyota hopes that the resulting "Large Behavior Model" will contain hundreds of tasks by the end of 2023, and over 1,000 by the end of 2024. Current experiments focus on kitchen tasks such as peeling potatoes or evenly spreading butter on bread.

In a related development, researchers at Google have developed PaLM‑E—a new kind of large language model (LLM) for controlling robots. Unlike previous attempts to use LLMs for robotic control, PaLM‑E can ingest raw streams of robot sensor data, as well as text data. Simpler projects integrate current LLMs with robotics. In one example, Microsoft used ChatGPT for control of robotic arms. In another, researchers at Heriot-Watt University and Alana AI created FurChat, a humanoid robot head that combines speech synthesis, synthetic facial expressions, and generative AI to hold spoken conversations with human partners.

Overall, signs suggest that humanoid robot technology will advance enough to create useful products within the next few years. And, investments in manufacturing plants should drive down costs if and when volumes increase. The bigger uncertainty is demand.

As Figure acknowledges on its website, two approaches to robotics applications exist—one is fitting the robot to the environment (humanoids) and one is fitting the environment to the robot (such as robot-optimized warehouses and factories). So far, the latter has proved more successful than the former. Robot-optimized environments not only simplify robot tasks, they sometimes have advantages over human-optimized ones. For example, robot-optimized warehouses or factories can make better use of 3D space. But even if the optimized-environment approach wins out for warehouses, factories, industrial spaces, and some other environments, it cannot work for the many human-occupied spaces where robots could work—but still generally don't. Homes, battlefields, retail stores, hospitals, hotels, and city streets cannot be reconfigured to exclude humans—and all are potential markets for humanoid robots.