Industry Insight: Humanoid Robots – Dexterous Hands: Tesla’s Upgraded Design and the Critical Role of Sensing

A dexterous hand must offer strong load capacity, tactile perception, high motion control precision, and multi-DOF mobility to complete diverse and complex operations. As such, the dexterous hand is a high-barrier technology and a central element in humanoid robot design.

Tesla has upgraded its humanoid robot Optimus with a 22-DOF dexterous hand—nearly doubling the previous generation’s 11 DOF. This significant leap allows for more refined motion and manipulation.

Based on the 2024 patent filing, Tesla’s Gen-3 hand likely adopts a tendon-driven architecture featuring planetary gearboxes, ball screws, and cable tendons. Motors and screws are embedded in the forearm. The motor rotation is converted into linear motion through the screw, which then pulls the tendon connected to the finger bones. To offset tendon deformation caused by wrist movement, stiff springs are wrapped around the tendons, similar to the mechanism in bicycle brake cables.

Unitree Robotics is one of China’s leading humanoid robot developers. Its robots show strong motion control and balance capabilities, although its dexterous hand still trails behind Tesla’s Optimus in terms of precision and design.

Recently, Unitree launched an in-house dexterous hand R&D project and has been actively hiring AI algorithm engineers, mechanical engineers, and full-stack mechatronics experts. This highlights their strategic focus on improving robotic hand functionality as a key competitive edge.

Sensor integration plays a central role in enabling dexterous robotic manipulation. According to Ziwei Xia et al., sensors for dexterous hands are categorized into internal and external sensors. Sensory processing operates at three levels: planning, control, and learning.

Internal sensors monitor joint positions, forces, and torques, providing real-time feedback to reduce errors and enhance motion accuracy. They are essential for precise actuation in dynamic environments.

External sensors include proximity, tactile, and multimodal sensors. Proximity sensors help estimate the distance between the hand and target object during the approach phase, while tactile sensors provide physical feedback about contact forces and object properties upon touch.

Early-stage tactile sensors focused primarily on force and pressure measurement. These are commonly embedded in fingertips and use piezoelectric effects to detect multidimensional forces. However, they often lack spatial resolution and struggle with detecting contact location.

To overcome this, multi-array tactile sensors have become increasingly common. Yet, most of these are single-modality—they can detect pressure but not temperature, texture, or other environmental factors.

That’s where electronic skin comes in. Multimodal sensors—capable of capturing a wider range of data types—can enable robots to sense force, temperature, texture, and more, mimicking human skin. As a result, electronic skin is expected to accelerate real-world deployment of dexterous robotic systems across diverse industries.