Five Fingers and Two Legs: Why Humanoid Robots Are Not Ready

"We are not going to have five-fingered hands in 20 years, and we will be lucky to have three-fingered hands." That verdict came from Rodney Brooks, co-founder and CTO of Robust AI, at RoboBusiness 2026, and it frames the gap between humanoid hype and engineering reality.

The annual robotics conference ran in Santa Clara, California, on the 20th and 21st of last month. Tesla has discontinued the Model S and Model X to concentrate on its Optimus humanoid. Unitree, Figure AI and Boston Dynamics are racing for early leads. Yet the machines still malfunction so often that deployment remains limited to research, exhibitions and controlled factory tasks. Ninety-eight years have passed since Eric, Britain's first robot, appeared in 1928.

The first day's keynote featured Brooks, UC Berkeley professor Ken Goldberg and Agility Robotics co-founder Jonathan Hurst discussing the next 20 years of robotics. Day two brought PSYONIC founder Aadeel Akhtar, Apptronik senior product manager Parker Conroy, Persona AI's Michael Patrick and Agility CTO Pras Velagudi. Five problems dominated.

Why are fingers still so hard?

A human hand has 27 bones, 27 joints and 34 muscles. Engineers judge that a humanoid hand needs 20 to 30 degrees of freedom to match it. Today's front-runners deploy 22 degrees of freedom in the fingers, 25 including the wrist.

The numbers alone do not close the gap. Robots still cannot adjust grip force to an object's weight the way a person does, and machines that grip stationary objects fail on moving ones because fingers, arms and eyes must work in concert. Even Elon Musk has marveled at how difficult fingers are. Brooks added: "There are no good three-fingered hands even in the lab yet. It will take more time before they are deployed."

A counter-argument is gaining ground: more fingers do not mean better performance. Agility Robotics uses two-fingered grippers on Digit 4 and Digit 5, citing a two-time Cybathlon winner who beat five-fingered competitors with a single-degree-of-freedom rotating parallel gripper. Goldberg said: "You can do a lot with a gripper. I think surgical robots are the best evidence that precise motions are possible with a simple gripper."

Wheels or legs?

Wheels are stable; legs carry a toppling risk. Chinese makers pushed bipedal development, and China's humanoid footraces and boxing matches sent ripples through the market. But bipedal balance still falls short of human ability, especially on inclines, uneven slopes and obstacle-strewn routes, so industrial sites must be perfectly controlled to prevent errors.

Apptronik, maker of the Apollo humanoid, hedges with both. Conroy said: "Many customers want to start with the wheeled version and expand to the legged one. There is also plenty of work only a legged robot can do." The sensors and upper body stay the same across both product lines.

Agility spent 10 years on "bird legs" — an ostrich-inspired design dating back to its Cassie robot — then switched Digit 5 to a forward-bending human-style leg, judging it better for lifting heavy objects and climbing.

"The problem we face now is not simple locomotion itself but local manipulation," Velagapudi said. He cited force disturbances when grasping objects, and pointed to mining, offshore oil drilling and shipboard work as markets that open up if locomotion is fully solved.

How do you train a physical machine?

Data collection may be the biggest obstacle. "The biggest difference between Optimus and cars is that with Optimus you cannot collect the vast amount of training data from a huge fleet of vehicles the way you can with cars," Musk said on last quarter's earnings call. Tesla is hiring Optimus data collection operators and building "Optimus Academy" for task practice.

Collection routes diverge sharply:

• Human Archive, an AI data startup, pays Indian laborers to wear head-mounted cameras and generate first-person work data.
• PSYONIC uses smartphones and smart glasses, syncing data with Nvidia training software, and partners with British device maker Koalaa to learn from prosthetic-hand patients. "Worldwide, as many as 10 million people have lost a hand, and 80% of them are in developing countries such as Pakistan and India," Akhtar said.
• Apptronik teleoperates many robots at once. Conroy called teleoperation data "the hardest data to collect" but "the most valuable," because the robot performs the task for real.

Quality beats volume, Velagapudi argued: "It is not about collecting a vast volume of data but about working out which data will be most useful."

What about safety and expectations?

No unified safety standard exists yet. The US relies on the private ANSI/RIA R15.08 standard for mobile robots, which Conroy called limited; extending ISO workcell guidance 108 globally matters, he said, along with proving safety to customers, insurers and bystanders through light and sound signals.

Digit 5 was designed to work outside cages, requiring safe perception, trustworthy decision infrastructure built with Nvidia, and motion constraints to limit risk.

Finally, expectations outrun hardware. "Every customer expects a robot to run 30,000 hours without failure, and Chinese robot hands are said to run for about 200 hours," Brooks said. Hurst put it plainly: "Humanoids went through a period of hype. Rather than trying to solve everything at once, you have to start by delivering small slices of value."