Xiaomi has scaled the CyberOne robot hand to human size, solving the heat dissipation issue through "sweating."

Title

Xiaomi shrinks CyberOne into a human-hand-sized mechanical arm by “sweating” to solve its heat dissipation problem

Abstract

This time, Xiaomi has modified its CyberOne mechanical arm—not by chasing algorithmic benchmarks and speed tests, but by focusing on real issues in the factory:

• Cutting volume by about 60%, bringing it close to hand size; 22 to 27 degrees of freedom; tactile coverage of more than 8,200 square millimeters on the palm
• Biomimetic sweat glands can dissipate 10 watts of heat, supporting continuous operation; a durability test ran more than 150,000 grasping-and-gripping cycles
• A factory task of tightening nuts achieves a 90.2% success rate; the training path follows a three-step process: tactile glove data, imitation learning, and reinforcement learning
• Open-sourcing TacRefineNet and 61 hours of tactile data; other teams’ data collection and model iteration can be much faster

In plain terms: Xiaomi isn’t rolling over model benchmarks—it’s tackling those unglamorous, bottleneck engineering problems: thermal management, durability, and tactile feedback.

Analysis

People working on humanoid robots and collaborative robots are increasingly coming to agree on a few things:

• Heat dissipation: Running for long periods leads to heat buildup—either reduce frequency or shut down. Sweat gland cooling may not sound flashy, but it directly determines whether a shift can be completed and whether production capacity can stay stable.
• Tactile coverage: With a tactile sensor covering 8,200 square millimeters on the palm, the robot can keep its grasping steady when vision is unclear or obstructed, without relying entirely on vision.
• Durability: 150,000 grasping cycles plus a 10-watt heat dissipation limit means you’re working at the boundary of mechanics and thermal science, not just stacking up paper parameters.
• Training approach: Combining tactile glove labeling, imitation learning, and reinforcement learning matches the real-world engineering workflow from demonstrations to online optimization.
• Industry direction: Tesla and Figure are also aggressively working on end-effectors, which shows everyone has recognized one key fact—the core pain point of sim-to-real isn’t only in the model; it’s also whether the hardware can truly handle it and whether the feedback quality is good enough.

The table below summarizes several key issues and Xiaomi’s solutions:

Things to note:

• 90.2% may be enough for certain stations, but for processes with high requirements for throughput rate it’s still a bit short. Whether it can be improved depends on whether the tactile-control closed loop and the durability strategy can be iterated together.
• The open-source TacRefineNet and 61 hours of tactile data still require adaptation when moved to non-Xiaomi hardware; they aren’t plug-and-play.

Impact assessment

• Importance: High (directly targets bottlenecks for industrial adoption)
• Type: Product launch, technical progress, industry direction

Judgment: For people watching the robotics and automation track, this is an engineering turning point that’s “slightly early on timing but clear in direction.” The real beneficiaries are teams with coordinated capabilities across hardware, tactile sensing, and control—plus long-term capital positioned across the industrial robotics supply chain. There doesn’t seem to be much opportunity for short-term trading.