On this page
The humanoid hand
Dolphins are intelligent. They communicate through signature whistles unique to each individual, coordinate hunts using complex vocalizations, and navigate...
Dolphins are intelligent. They communicate through signature whistles unique to each individual, coordinate hunts using complex vocalizations, and navigate their world through sophisticated echolocation. Some biologists argue they may rival us in raw intelligence. Yet dolphins never built cities, tools, or technology. Not because they lacked imagination or the ability to communicate, but because they lacked hands.
Without hands, intelligence has no leverage. Dolphins could double their brainpower, develop even more intricate languages, or dream of stars in the deep. None of it would matter. They cannot strike a spark, weave a net, or carve stone. Communication and intelligence alone are not enough. Without the ability to manipulate the physical world, intelligence remains trapped inside the mind.
The hand is what unlocked human civilization. Language became writing, ideas became machines, thought became progress. All because of ten fingers and two opposable thumbs.
This is why humanoid robots need more than legs. Every object we have created, from a doorknob to a smartphone, assumes the presence of a human hand. A humanoid without hands is a half-finished machine.
Entire companies now exist solely to replicate the human hand, aiming for hands that are not only functional but sensitive, adaptive, and capable of the same fluid dexterity we take for granted.
The 2025 Hand Race#
At Tesla's "We, Robot" event in October 2024, held at Warner Bros. Studios in Burbank, California, attendees noticed something remarkable on display: a new hand for Optimus. Though the humanoid robots serving drinks and mingling with guests were still Gen 2 models, the hand on the pedestal represented the future. It was tendon-driven, mimicking the way human hands actually work, and promised to double the dexterity of earlier versions.
The full reveal came six weeks later. On November 28, 2024, Tesla released a video showing the new Gen 3 hand in action. The upgraded design featured 22 degrees of freedom, up from just 11 in earlier versions, plus 3 additional degrees in the wrist and forearm. A video showed the hand catching a tennis ball mid-air, its fingers flexing and adjusting with uncanny precision. Milan Kovac, head of Tesla's Optimus program, called it a "milestone achievement."
Elon Musk had been blunt about the stakes. In internal meetings, he revealed that more than half of the Optimus Gen 3 engineering effort focused on the hands. Locomotion was solvable. Manipulation was the bottleneck.
One month earlier, a smaller Chinese competitor had made its own move. XPeng's Iron humanoid robot revealed hands featuring what the company called "the world's tiniest harmonic joint," also with 22 degrees of freedom. The convergence was no coincidence.
By late 2025, every serious humanoid company had arrived at the same conclusion. The hand is the bottleneck. You can build a robot that walks flawlessly, navigates autonomously, and reasons through complex tasks, but if it cannot manipulate the physical world with human-level dexterity, it remains a demonstration rather than a product.
This triggered a global sprint. Figure AI equipped its Figure 02 with hands boasting 16 degrees of freedom and the ability to carry objects up to 25 kilograms. By October 2025, the company had redesigned the entire hand system for Figure 03, prioritizing what it called "purpose-built manipulation" for Helix, its vision-language-action AI. The breakthrough went beyond mechanics: Figure 03 embedded cameras directly into the palms of its hands, giving the robot a close-up, first-person view of manipulation tasks. This solved a critical occlusion problem where the robot's own fingers would block the view from head-mounted cameras. Combined with custom tactile sensors capable of detecting forces as small as 3 grams, Figure 03's hands represented a new class of manipulation tools.
Meanwhile, in Norway, 1X Technologies took a different path. Instead of chasing maximum degrees of freedom, the company bet on compliance and safety. Its NEO humanoid, launched for pre-order in October 2025 at $20,000, featured 22-degree-of-freedom hands powered by a patented Tendon Drive system. Unlike the rigid actuators used by competitors, 1X's approach used high-torque-density motors to pull synthetic tendons, creating movements that were gentle, compliant, and eerily quiet. At just 22 decibels, NEO's operation was softer than a modern refrigerator. The tendon-driven design prioritized safety for home deployment, where robots would operate around children and pets without constant supervision.
The tendon-driven approach was not new. Shadow Robot Company in London had pioneered it decades earlier with the Shadow Dexterous Hand, a research-grade marvel featuring 24 joints and 20 degrees of freedom. But Shadow's hands were too expensive and fragile for commercial use. The original version cost upwards of $100,000, and while a newer model launched in 2024 started at $74,000, it remained confined to laboratories and AI research. Tesla, 1X, and others were now racing to bring tendon-driven dexterity to mass production.
Even China's budget player, Unitree, recognized the stakes. Its G1 humanoid, launched at just $16,000, came equipped with optional 3-fingered force-controlled dexterous hands. The hands were simpler than those of Western competitors, but Unitree's bet was on volume and iteration. Ship units, gather data, improve fast.
The Challenge of Copying Evolution#
Replicating the hand is among the hardest problems in robotics. Evolution spent millions of years refining it.
The human hand contains 27 bones, 34 muscles, over 100 ligaments, and a dense network of nerves that can detect a grain of sand under a fingertip. Humans switch effortlessly between raw force and delicate precision: lifting a heavy suitcase, then peeling a grape without tearing it.
Robots struggle with this duality. A gripper strong enough to lift a drill may crush fragile objects. A delicate manipulator that can pick up a strawberry may fail at twisting a wrench. Building fingers that move is straightforward. Building fingers that adapt, sense, and adjust in real time remains an open engineering challenge.
Most humanoids today rely on simplified hands. Many are little more than clamps with symbolic fingers. Enough to hold boxes, not enough to replace the subtleties of human touch.
Competing Philosophies#
Roboticists divide along design philosophies, each shaped by different priorities.
Some build rigid mechanical hands: precise, powerful, but heavy and expensive. Others experiment with soft robotics, using flexible materials that mimic muscles. These are safer and more adaptable, but often too weak for heavy lifting.
Another approach is underactuation. Instead of controlling each finger joint independently, a few motors pull on tendons that move multiple fingers at once. This reduces cost and complexity but sacrifices the independence that human hands have. Tesla's Optimus and 1X's NEO both use variations of this approach, but with different goals. Tesla prioritizes manufacturability and speed, aiming for hands that can be mass-produced in its Fremont factory. 1X prioritizes safety and compliance, designing for a robot that will operate unsupervised in homes alongside children and pets.
The frontier lies in sensor-rich hands. Outfitted with tactile sensors and force detectors, these hands can feel the texture of a surface or the resistance of a grip. They give robots something like a sense of touch, but processing this information in real time requires sophisticated AI.
The closer a hand gets to human-level dexterity, the higher the price and fragility. Research labs build astonishing prototypes. Commercial humanoids compromise with simpler designs that can ship and scale.
Milestones in Robotic Hands#
Several legendary hands have demonstrated what was possible, even if they never reached mass production.
The Shadow Dexterous Hand, developed in London, remains one of the most sophisticated. With 24 joints and tendon-driven mechanics, it is so precise that AI researchers use it to train reinforcement learning systems. Google DeepMind collaborated with Shadow to develop a newer version called DEX-EE, built specifically for machine learning research. The company's latest model, launched in 2024, starts at $74,000 depending on configuration.
Germany's DLR Hand pioneered tendon-driven mechanics decades ago, demonstrating extremely fine manipulation in laboratory settings.
NASA's Robonaut hand was designed for space missions, capable of gripping tools that astronauts use in the harsh environment beyond Earth's atmosphere.
These hands remain niche, too expensive and fragile for daily use in a factory or a home.
Modern humanoid companies have taken more pragmatic routes. Tesla's Optimus hand is optimized for cost and manufacturability rather than perfect dexterity. Figure AI's humanoid uses hands with embedded palm cameras and tactile sensors, deployed at BMW's South Carolina plant for car manufacturing tasks. Unitree ships its G1 humanoid with optional 3-fingered hands that are closer to functional tools than anthropomorphic replicas, but at $16,000 for the base model, the company has made humanoids accessible to universities and small research labs for the first time.
At one end of the spectrum, expensive research platforms like the Shadow Hand. At the other, mass-produced but limited hands that prioritize affordability. In 2025, the gap between the two is finally closing.
Why Touch Changes Everything#
Without tactile sensors, robots must guess grip strength. You do not crush a paper cup because your nerves instantly tell you how much pressure you are applying. You do not drop a glass because your fingertips detect its slip and adjust your grip unconsciously.
Robots lack this reflex. This is why AI researchers treat dexterity as a benchmark for general intelligence. Training a robotic hand to rotate an object, thread a needle, or stack blocks forces the system to integrate perception, planning, and action in real time.
Progress in tactile sensing is moving fast. Electronic skins are being developed with thousands of tiny sensors that detect pressure, vibration, and temperature. Combined with AI motor control, these give humanoids something close to a sense of touch. Shadow Robot's BioTac sensors, available for its hands, can detect micro-vibrations and temperature gradients. Unitree's G1 offers optional tactile sensor arrays. Figure 03's hands feature custom tactile sensors that can detect forces as small as 3 grams. 1X's NEO uses its Tendon Drive system to provide inherent force feedback, allowing the robot to sense resistance through the tension in its synthetic tendons.
Figure 03 added something beyond tactile sensing: vision in the hands. By embedding cameras directly into the palms, the robot gains a close-up view of exactly what it is manipulating. When fingers curl around an object, the palm cameras see what the head-mounted cameras cannot. This extra sensory layer, combining vision with touch, brings humanoid hands closer to the multimodal sensing that humans take for granted.
More Than a Tool#
The hand is also social. We shake hands to seal trust, gesture to emphasize meaning, wave, point, console.
A humanoid that cannot gesture is incomplete in environments where physical interaction matters. In hospitals, homes, and schools, humanoids will not only manipulate objects but also interact with humans. A hand extended in greeting, an open palm of reassurance, these matter as much as the ability to hold a screwdriver.
When 1X unveiled NEO in October 2025, the company emphasized that the robot was designed to "complement living spaces rather than disrupt them." The soft, compliant hands were not just about manipulation. They were about trust. A robot with rigid, claw-like grippers would never feel safe in a home. A robot with hands that moved gently, quietly, and naturally could become part of the household.
The Road Ahead#
Three converging breakthroughs are redefining the humanoid hand in 2025.
First, new materials. Soft polymers, synthetic muscles, and flexible actuators are making hands safer and more adaptable. Tesla's Gen 3 hand uses custom-designed actuators built in-house. 1X's Tendon Drive system uses what the company claims are "the highest-torque density motors on earth" to power tendon-based transmissions.
Second, multimodal sensing. Beyond traditional tactile sensors, robots are now gaining palm-embedded cameras for close-up vision and electronic skins with dense arrays of sensors. Figure 03's embedded palm cameras solve occlusion problems while its tactile sensors detect forces as small as 3 grams. Shadow Robot's latest DEX-EE hand features hundreds of taxels (tactile pixels) per finger, each capable of detecting minute changes in force and texture.
Third, AI motor control. Learning-based systems refine dexterity through millions of practice grasps, going beyond what engineers can hard-code. Tesla trains Optimus hands using its Full Self-Driving neural network architecture. Figure AI uses its Helix vision-language-action model to enable Figure 03 to learn manipulation tasks from human demonstrations. 1X's Redwood AI model, a 160-million-parameter transformer, combines vision, touch, and body movement data trained on real-world teleoperation.
Once these mature, humanoid hands will rival ours. They will chop vegetables, fold clothes, assist in surgery, repair machinery, and tie shoelaces. And when they do, the economics of humanoid labor will shift permanently.
References: