90% of energy is converted into waste heat! The cooling problem of humanoid robots is a "critical bottleneck" for commercialization.

As humanoid robots move closer to mass production, heat dissipation becomes a “hard barrier”—it’s not just about lowering temperatures but directly affects joint torque output, chip thermal throttling, and safe fast charging of batteries. Guohai Securities’ research on the mechanical industry elevates thermal management from a “supporting role” to the forefront, analyzing where humanoid robots generate heat, where bottlenecks occur, and possible engineering solutions.

Guohai Securities’ Chief Mechanical Analyst Zhang Yuying wrote in the report that “90% of the energy produced by humanoid robots is directly converted into heat, accumulating in small spaces such as motor windings, gearboxes, and chips,” and in extremely compact structures like dexterous hand joint chambers, “the gap inside may be less than 2mm,” making traditional 5mm centrifugal fans physically impossible to install.

The heat problem is more than just “hot.” It can trigger chip thermal throttling, causing efficiency to collapse; high temperatures also reduce signal transmission stability and can push the robot’s continuous operation capability to the edge of protective modes. The report extensively discusses copper loss, iron loss, wind abrasion in joint motors, as well as temperature constraints of drivers, reducers, and encoders, ultimately weighing options between air cooling, liquid cooling, and chip control.

More notably, heat dissipation isn’t limited to joints. The torso’s batteries and computing stacks are also heat sources. Patents from Tesla and Figure AI focus on “airflow pathways, intake and exhaust placement, and shared cooling interfaces for computers and batteries.”

Low energy efficiency and dexterous hands are the ultimate test for heat dissipation

The report compares “humanoid robots (estimated) vs. humans,” concluding that: at the same power level, robots have significantly lower energy conversion efficiency, making heat more prone to accumulate in tight spaces. Once heat builds up, the first issues often aren’t shell temperature but junction temperature of chips and drive losses.

The report describes a typical positive feedback loop: after thermal throttling is triggered, the RDS(ON) of drive chips exhibits positive temperature characteristics; for every 10°C increase in junction temperature, resistance increases by about 4%, which further raises I²R losses, generating more heat, and causing system efficiency to “collapse” in a snowball effect.

Two other consequences are highlighted: electromagnetic interference and signal stability degrade in high-heat environments; and the robot’s sustained high-speed operation weakens, frequently entering protection mode, directly limiting application potential.

The heat dissipation challenge is especially amplified at dexterous hand joints: space is extremely limited, with the report noting gaps inside chambers may be less than 2mm. Traditional 5mm centrifugal fans cannot be installed—meaning mature air cooling solutions used in industrial equipment are ineffective here. Meanwhile, dexterous hands require high power density output, lightweight design, and small volume—three conflicting requirements.

The motor layout in dexterous hands also impacts heat dissipation. Current mainstream solutions include internal placement (motors inside the palm or fingers), external placement (drivers in the forearm), and hybrid configurations. The report suggests Tesla’s next-generation dexterous hand may adopt a hybrid layout: wrist-mounted motors combined with palm-in motors, driven by tendon cables. This design moves heat-generating motors to the larger wrist space, providing more room inside the fingers.

Copper loss is fundamentally a “trade-off” among volume, torque, and temperature rise

The report breaks down joint motor losses in detail. Typical proportions are: copper loss (stator windings) 40%-60%, iron loss (stator core, hysteresis + eddy currents) 20%-30%, mechanical losses (bearings/air gaps) 5%-10%, permanent magnet loss (rotor magnets) 5%-10%. In drive modules, power devices (MOSFETs, etc.) switching/conduction losses can account for 30%-60% of total drive loss.

All these losses ultimately manifest as “temperature red lines”: winding temperature <155°C; encoders <100°C-120°C; reducers possibly only <65°C (the report cites an example: if the rated temperature of a reducer is 65°C, and the motor winding temperature nearby can be at most 15°C higher, then the winding maxes out at 80°C, constraining motor design). Cooling isn’t just about making motors stronger; neighboring components like reducers, feedback devices, and bearings also set limits.

The report’s approach to copper loss management resembles a “combination of structure + materials + algorithms.” Under high dynamic conditions, compact modules have insufficient cooling area, leading to high natural convection thermal resistance; temperature gradients can cause asymmetric thermal deformation, resulting in multi-degree-of-freedom pose errors. Engineering thus becomes about balancing motor volume, torque, and heat.

Proposed directions include:

• Structure: develop high-heat-dissipation structures, improve heat transfer efficiency. For example, the report cites a patent for a flattened rotary joint module in the “New Sword” transmission, using a parallel structure of motor and reducer to alleviate heat buildup in compact designs.
• Materials: use low thermal expansion alloys for screw rods to reduce deformation; high thermal conductivity carbon fiber composites for housings in the “New Sword” scheme.
• Algorithms: real-time temperature compensation to reduce drift; use of heated sensors, and reducing current/torque/stopping when necessary.

Iron loss and rotor eddy currents: first suppress harmonics, then add cooling fins

Regarding iron loss, the focus is on rotor eddy current loss: high-speed permanent magnet rotors operate in complex magnetic fields, where harmonic magnetic fields cause induced voltages in conductive rotor parts, generating eddy currents and associated losses. The report emphasizes that underestimating eddy current loss can lead to rotor overheating and safety hazards, and flawed cooling design.

Mitigation strategies are summarized as two main approaches:

• Suppress asynchronous magnetic fields at the source: aim for sinusoidal air-gap magnetic fields via optimized stator tooth-slot structures, winding configurations, increased power device switching ratios, adding filtering reactors between motor and controller, and enhancing stator inductance from the motor design—each with trade-offs in size, weight, dynamic response, losses, and cost.
• Shield propagation paths: use shielding layers on casings to block alternating magnetic fields; or create slots/holes on casings to cut eddy current loops.

Air cooling remains the most economical, liquid cooling more effective

The report does not dismiss air cooling as “outdated,” instead emphasizing it remains an economical and reliable option: natural convection suits devices with heat flux densities below 0.8 W/cm²; forced air cooling can achieve 5-10 times the effectiveness of natural cooling. The challenge lies in structure: dexterous hand joint chambers with gaps less than 2mm make traditional 5mm fans impossible to install. The report provides two “fit” solutions:

• UBTECH’s joint patent: simplified joint module structure, hollow design, and external heat dissipation devices (like fans) for air cooling.
• MEMS micro-fans: thickness less than 1.5mm, can be embedded into hotspots like chips and motor drivers, using piezoelectric actuation to produce micro jets for targeted cooling; the report notes heat flux handling capacity over 100 W/cm².

Compared to air cooling, liquid cooling introduces cold plates, fluid circuits, pumps, and expansion tanks—more complex but with higher thermal conductivity and heat capacity, suitable for higher heat flux scenarios. The report lists common liquid cooling methods: circulation, immersion, spray cooling.

In the “oil cooling” section, it cites a “New Sword” planetary roller screw oil cooling scheme: hydraulic oil flows through channels and holes to form a static pressure film at contact surfaces, reducing friction and wear; it also carries away heat, reducing thermal deformation. Such solutions address both cooling and lifespan but introduce sealing, circuit, and maintenance challenges.

Thermal management isn’t just about structure and fluids; chip-level solutions also matter.

The report dedicates a section to “chip control,” emphasizing that better control can reduce drive current, naturally lowering heat. For example, Peak AI’s high-performance stepper motor control chips enable closed-loop operation at lower currents, improving reliability and reducing heat. The product lineup includes main control chips (MCUs/ASICs), driver chips (HVICs), and power devices (MOSFETs). The FU75xx series MCUs are used in robot joints and dexterous hands.

Batteries and computing in the torso: Tesla and Figure AI connect heat sources via airflow pathways

Battery section highlights a practical trade-off: robot battery applications balance “standardization vs. performance premium.” It notes two industry trends:

• LG Energy Solution: launching 2170 cylindrical cells for robots/drones at InterBattery 2026, with performance tiers; H52A supports up to 8C high-power fast charging in about 15 minutes.
• Xiaopeng: releasing humanoid robot IRON in November 2025, emphasizing all-solid-state batteries—30% weight reduction, 30% capacity increase, aiming for mass production by late 2026.

On patents, the report illustrates “torso thermal management” trends with Tesla and Figure AI examples:

• Tesla’s design involves energy storage device enclosures and computing systems forming duct pathways, with fans driving airflow through heat sinks for batteries and computers; multiple fans, vents, and heat sinks are configurable, with materials balancing structural support and heat dissipation.
• Figure AI’s design places inlets/outlets at the lower edge of the waist and underarm sides, with fans drawing fresh air into the upper torso, exchanging heat, and exhausting from lower torso/waist vents; airflow can cool computing devices during operation (GPU, CPU), and during charging, it cools batteries to shorten charging time and support higher currents. Their patent specifies battery capacities of 1.5–5 kWh (preferably 2–3 kWh), with a range of 2.5–8 hours (preferably at least 3.5 hours).

Risk warning and disclaimer

Market risks exist; investments should be cautious. This article does not constitute personal investment advice and does not consider individual user’s specific investment goals, financial situation, or needs. Users should determine whether any opinions, viewpoints, or conclusions herein are suitable for their circumstances. Investment based on this information is at their own risk.