Why Robot Hands May Be the Hardest Part of the Humanoid Race

The hand is where a humanoid meets the economy

A robot can navigate a warehouse and still create little value if it cannot handle the objects inside it. Most workplaces do not need a machine merely to stand near a shelf. They need it to grasp a package, insert a component, operate a tool, open a door or transfer an item without dropping it. Manipulation converts mobility into useful work.

Humanoid form is attractive because factories, offices and homes already contain objects designed for people. The same advantage creates a demanding specification. A human-compatible hand must reach controls and use tools without requiring the environment to be rebuilt, yet it must do so with mechanical and sensory systems that remain far less compact and adaptive than biology.

A hand contains many problems in very little space

Human hands combine multiple joints, muscles, tendons, soft tissue, skin and dense sensory feedback. A robotic equivalent must place actuators, transmissions, sensors, electronics and cables inside a similarly constrained volume. Adding another controllable joint can improve dexterity while increasing weight, wiring, calibration and potential failure points.

Shadow Robot’s research hand illustrates the scale of the design challenge: it specifies 24 degrees of freedom, 20 motors and more than 100 sensors. That platform is built for advanced research, not evidence that the same complexity is already economical for mass-produced humanoids. Commercial designers must choose how much dexterity is worth its cost, power consumption and maintenance burden.

Degrees of freedom create both possibility and complexity

A simple parallel gripper may have one commanded opening. It is predictable, strong and easy to model, but cannot form the varied grasps needed for scissors, cloth, a key or a small pill. More joints allow fingers to wrap, pinch, reposition and conform to different objects. They also expand the control space.

The system must decide the position and force of each joint while coordinating the wrist, arm and sometimes the robot’s balance. Several joint configurations may achieve the same visible grasp, but only some will remain stable as the object moves. Mechanical coupling or underactuation can reduce the number of motors, although it also reduces independent control. There is no free increase in dexterity.

Touch matters when vision becomes unreliable

Cameras can estimate an object’s position before contact. Once fingers close around it, the hand may block the view. Transparent, reflective, deformable or visually similar objects can also confuse perception. Tactile sensing reveals contact location, pressure, vibration or shear—the sideways force associated with an object beginning to slip.

Shadow Robot’s tactile work emphasizes that sensors must conform to curved fingertips and survive as part of the contact material. Figure’s 2026 Helix 02 demonstration describes the use of vision, touch and proprioception for tasks including separating small objects and handling a syringe. These examples show why tactile data is becoming central. They do not establish performance across every material, contamination level or operating duration.

Force must be strong and gentle at the same time

A useful hand may need to hold a heavy tool and then pick up a fragile object. Position alone is insufficient: the same closing motion can secure one item and crush another. The controller must estimate friction, compliance and object geometry, then apply enough force to prevent slip without causing damage.

Contact can change in milliseconds. A delayed cloud decision cannot correct every small slip, so low-level control must run locally at high frequency. Learned policies may choose a grasp or adapt it, while fast controllers stabilize the fingers. The division resembles whole-body robotics: slower reasoning sets the objective and faster control preserves physical stability.

Softness helps safety but complicates precision

Compliant fingertips and coverings can spread pressure, improve friction and reduce injury at contact. They also deform, wear and change the relationship between motor movement and fingertip position. A rigid mechanism is easier to model precisely; a compliant one can tolerate uncertainty but requires the controller to understand that deformation.

Materials must resist cuts, oils, dust and repeated cleaning. A surface optimized for gripping cardboard may behave differently on wet glass or fabric. Replaceable fingertip coverings can improve serviceability, but seams and connectors introduce their own failure modes. The correct material depends on the task environment rather than visual resemblance to skin.

In-hand manipulation is harder than picking and placing

Lifting an object with a stable grasp is only one level of manipulation. Humans rotate a pen, shift a key toward the fingertips or reposition a screw without putting it down. This in-hand manipulation requires controlled changes in contact while preserving the object against gravity and unexpected motion.

Demonstrations often begin with successful pickup and end after placement. Industrial value may require orientation, insertion and tool use, where millimetres and force direction matter. A robot that can grasp many objects but cannot reliably align them may still need specialized fixtures. Dexterity should therefore be evaluated through the complete task, not the most photogenic moment.

The hand cannot be separated from the arm and body

A grasp depends on approach angle, wrist pose, arm compliance and balance. When a humanoid pulls a drawer or lifts a large object, the force travels through the entire body. Treating the hand as an isolated endpoint can produce a locally stable grasp that destabilizes the robot.

Figure’s Helix 02 announcement is notable because it presents walking, manipulation and balance under a unified visuomotor system. Whole-body coordination is the relevant direction for humanoids working in unconstrained spaces. The reported four-minute dishwasher task is a company demonstration and should be understood as evidence of capability, not yet a general measure of fleet reliability.

Training data for dexterity is expensive

Language models can learn from enormous collections of digital text. Robot-hand data requires physical interaction or a useful simulation. Teleoperation can capture human demonstrations, but mapping a human hand to different robot geometry is difficult, and collecting diverse failures takes time. Each hardware revision may also change the data distribution.

Simulation can generate experience quickly and safely, although friction, flexible objects and detailed contact are difficult to reproduce exactly. The gap between simulation and reality becomes especially visible in fingertips, where small errors alter whether an object slips. Human video offers scale but does not directly contain force and tactile measurements. Practical training will combine sources rather than rely on one universal dataset.

Recovery is a better test than a perfect demonstration

Real objects move, deform and arrive in unexpected orientations. A hand must recognize a weak grasp, set an item down safely and try again. Figure’s laundry demonstration reports recovery from picking multiple towels, an example of why error handling matters. A system that completes a staged task only when every grasp succeeds is not ready for an uncontrolled environment.

Evaluation should count drops, damage, interventions, retries and recovery time across many cycles. Success rate alone can hide cost: 95 percent may sound strong, but one failure in twenty actions can stop a high-volume line. The acceptable rate depends on consequence and whether the process can isolate failures without human rescue.

Durability may be harder than a laboratory benchmark

Fingers are exposed to collision and concentrate force in small components. Tendons can stretch, gears can develop backlash, sensors can drift and soft coverings can tear. A hand that performs a delicate task once may not survive thousands of cycles with the same calibration.

Research and commercial requirements therefore diverge. A research hand maximizes observability and dexterity for experiments. A production hand must balance performance with mean time between failures, replacement time, part availability and field calibration. Figure’s redesign of Figure 03 around manufacturing and its new hand system reflects this wider problem, but claimed readiness for scale must ultimately be tested through production and deployment data.

Cost decides whether human form is justified

More actuators and sensors add component cost, assembly labour, compute and maintenance. The hand must enable enough additional tasks to justify that expense. In a fixed process handling uniform boxes, a two-finger gripper can be faster, stronger and easier to certify. Anthropomorphic dexterity becomes valuable when the object set and environment vary.

This produces a spectrum rather than one winning design. Some humanoids will use simplified hands optimized for logistics. Others may need tactile fingertips and individually controlled fingers for homes, laboratories or skilled tool use. Modular end effectors can outperform a permanent human-like hand when task changes are predictable.

What buyers and observers should ask

Count complete tasks and operating cycles, not only degrees of freedom. Ask which objects were unseen, whether the sequence was autonomous, how many attempts failed, whether touch was used, how long calibration lasted and what happened after a slip. Payload and closing force matter, but so do delicacy, speed and repeatability.

Also distinguish teleoperation, learned autonomy and scripted motion. Teleoperation can be commercially useful and can generate training data, but it is not the same as autonomous manipulation. A video edited from successful attempts does not provide a fleet-level rate. Credible reporting states the test conditions and unresolved limitations.

What is likely to improve next

Tactile sensors should become more integrated, durable and easier to manufacture. Learned controllers will combine vision, touch and proprioception at higher frequency. Better human demonstration, simulation and fleet data may improve generalization, while compliant mechanisms can reduce the need for exact geometric models.

Progress will not make every hand human-equivalent at once. Advances may first produce reliable task families: parcel handling, laundry, simple tool use or selected assembly. The important signal will be sustained performance across object variation and thousands of cycles, not a new isolated feat.

The NTS View

Hands may be the hardest part of the humanoid race because they compress nearly every robotics challenge into the point of contact. The mechanism must be compact yet strong, compliant yet precise, sensitive yet durable. The software must combine uncertain perception with fast physical control and recover when the world does not match the training data.

Recent demonstrations show real progress in tactile sensing, learned finger control and whole-body manipulation. They should be evaluated as milestones rather than universal solutions. The gap between picking an object once and manipulating varied objects reliably, economically and safely remains large.

The winners may not build the hand that looks most human. They will build the level of dexterity their target work actually requires, prove it across long operating cycles and make repair part of the design. In robotics, usefulness is measured not by how many fingers move, but by how consistently those fingers complete the job.