Dexterous Manipulation

Dexterous manipulation is the source’s term space for high-degree-of-freedom hands, whole-body control, and fine object handling in Embodied AI. 170: 【具身季报 26Q2】世界模型大风不停，和不想被贴标签的人 treats dexterous hands as a key infrastructure layer because hand structure, degrees of freedom, drive method, sensors, thermal behavior, and retargeting constraints shape what robot data can be collected and what policies can transfer.

The episode contrasts direct-drive hands, cable-driven hands, and hybrid approaches. It presents 5G Robotics as a possible infrastructure supplier for research, Genesis Robotics as a model/data company demonstrating early dexterous tasks, and Tesla as an industrial route whose Optimus hand choices can pull other large engineering teams toward cable-driven designs.

从会跳舞到有感知，触觉是机器人通往智能的门票吗？｜ S10E19 adds Tactile Sensing as the missing feedback layer for manipulation. Eric Li Zhiqiang / 李志强 argues that preprogrammed demos such as grabbing eggs or peeling shells do not prove general dexterity unless the robot can detect contact, force, slip, softness, texture, and error in real time across new objects.

Key Claims

• Deformable objects such as clothes, soft packages, food, and plastic bags expose limits in rigid-body assumptions and simple grippers.
• High-DOF hands are not interchangeable data devices; different hand geometry, motors, and sensors make retargeting difficult.
• Independent hand suppliers may control an important data gateway if their hardware becomes widely used; full-stack robot-body companies may pull that data layer back inside the body.
• Fine insertion, fastening, gripping, and medical puncture tasks often depend on “feel” that is hard to specify visually or linguistically.
• Optical Tactile Sensing is presented as one route for giving dexterous hands dense contact feedback without treating touch as ordinary vision.

Connections

• 5G Robotics, Genesis Robotics, Tesla, and Unitree Robotics — hardware and ecosystem examples in the episode.
• Robot Logistics Sorting — practical scenario where dexterity matters.
• Embodied Robot Data Paradigms and Real Robot Data Strategy — data implications of hand choice and retargeting.
• Tactile Sensing, Optical Tactile Sensing, Yimu Technology / 一目科技, and Tactile Transformer Encoder — touch, hardware, company, and model-interface additions from the What’s Next source.