Dexterous Manipulation

Dexterous manipulation is a robot's ability to grasp, reposition, and use objects with the precision, adaptability, and force sensitivity of a human hand. It is widely considered the hardest unsolved problem in robotics — and the capability that separates a general-purpose humanoid robot from an expensive bipedal novelty. A robot that can walk but not fold laundry, use a screwdriver, or handle an egg without crushing it has limited commercial value. Dexterous manipulation is where the value lives.

Why Hands Are Hard

The human hand has 27 bones, 34 muscles, and over 17,000 tactile receptors. It can thread a needle, crush a walnut, and type on a keyboard — tasks spanning force ranges from millinewtons to hundreds of newtons, with sub-millimeter position accuracy. Replicating this in hardware is a mechanical engineering challenge; replicating the control is an AI challenge of a different order. Manipulation involves continuous contact dynamics (how surfaces slide, stick, and deform), real-time force estimation (how hard am I squeezing?), and planning under uncertainty (will this object slip if I tilt it?). These physics are among the hardest to simulate, making sim-to-real transfer for manipulation more challenging than for locomotion.

The State of the Art in 2026

Tactile sensing has emerged as a breakthrough enabler. Sharpa, in collaboration with NVIDIA, developed tactile sensing systems that give robot fingers human-like touch sensitivity, enabling them to detect slip, estimate object hardness, and adjust grip force in real time. This addresses a fundamental limitation of vision-only manipulation: you can't see how hard you're squeezing, and you can't see the moment an object starts to slip.

In-hand manipulation — repositioning an object within the hand without setting it down — has advanced significantly. Recent research demonstrates robots manipulating articulated tools (scissors, pliers) in-hand, using controllers that combine simulation-trained base policies with sensor-driven refinement learned from hardware demonstrations. Physical Intelligence has demonstrated laundry folding, a task requiring continuous deformable-object manipulation that was considered intractable just two years ago.

NVIDIA GR00T N1.7 shipped in March 2026 with advanced dexterous control as a standard capability, marking the first commercially licensed foundation model with general-purpose manipulation skills.

Gripper Design

Robot hands span a spectrum from simple parallel-jaw grippers (two flat fingers, highly reliable, limited dexterity) to anthropomorphic five-finger hands (maximum dexterity, maximum complexity). Most commercial deployments in 2026 use a middle ground: two- or three-finger adaptive grippers with compliant (slightly soft) fingertips that conform to object shapes. Fully anthropomorphic hands remain largely in research, though companies like Shadow Robot and several Chinese firms are pushing five-finger designs toward commercial readiness. The design choice reflects a tradeoff: more fingers enable more tasks, but each additional degree of freedom multiplies the control complexity.

The Remaining Gap

Despite progress, the gap between human and robot manipulation remains large for deformable objects (cloth, rope, food), multi-step assembly (putting together furniture, connecting cables), and tool use in unstructured environments. These tasks require reasoning about physics, planning multi-step sequences, and recovering from errors — capabilities that connect manipulation to VLA models and world models. The hand is the effector, but the brain decides what to do with it. Closing this gap is the central research and engineering challenge for humanoid robots in the next 3–5 years.