On November 2, the inaugural CMG World Robot Mountain Climbing Competition kicked off in Tai'an. On the winding path from Hongmen Gate to Zhongtian Gate on Mount Tai, a Unitree G1 crouched by the stone steps, head down. The operator explained that the robot had overheated. Subsequently, staff carried it down the mountain on a stretcher.
This G1 is the "Spring Festival Gala model" that performed martial arts on stage. Standing approximately 132 cm tall and weighing 35 kg, it features a 9000 mAh battery with a nominal runtime of about two hours. Its knee joints deliver a maximum torque of 120 N·m, with a retail price starting at CNY 99,000.
In an exhibition hall, performing a few high-difficulty moves is a breeze for it; running on flat ground poses no problem either. However, the challenge presented by the Mount Tai competition was 3.9 km, over 3,100 steps, and an elevation gain of approximately 600 meters. The stone steps were slippery, with the narrowest sections measuring only 13 cm.
During continuous ascent, the joint motors sustained high-torque output, causing heat to accumulate until it reached the protection threshold, triggering active system throttling and shutdown. The operator stated, "We would rather carry it down than allow overheating to damage servo lifespan."
The G1 was not the only robot to stumble at the competition site. On the same day's course, some robots barely finished while being pulled by ropes around their necks, others collapsed at shop entrances and were removed by firefighters, and some teams even laid their robots flat before the race to save power and cool them down. The G1 was simply the one captured most completely by cameras.
We are not here to dampen enthusiasm, but to examine the costs hidden behind the stretcher—costs that remain invisible within the exhibition hall.
01、Heat Dissipation Is Not Just About "Adding a Fan"
Inside humanoid robot joints reside high-power motors, driver chips, and reducers. A complete unit integrates 20 to 50 actuators, with joint actuators contributing 50% to 60% of the total heat generation. High-load chips in computing units can reach instantaneous power consumption of 50 to 300 W, resulting in extremely high heat flux density. Power conversion and driving devices in the power system add another 12% to 15% of thermal losses.
These heat sources are not evenly distributed but concentrated in specific hotspot regions. If hotspots are not properly managed, the issue is not merely "running hotter," but leads directly to frequency throttling, torque limitation, or even protective shutdown.
More critically, Unitree's official website explicitly lists "localized air cooling" for the G1's thermal management system. Air cooling is mature, cheap, and easy to maintain, but its effectiveness for high-power joints has inherent limits: noise, dust ingress, and restricted airflow paths. In scenarios like Mount Tai involving sustained full-load climbing, localized air cooling is inevitably pushed to its breaking point. Liquid cooling offers 10 to 50 times higher thermal efficiency than air cooling and can reduce core temperatures from 100°C to below 60°C. However, liquid cooling introduces tubing, pumps, seals, maintenance requirements, and additional weight. For a lightweight humanoid robot weighing 35 kg with a price tag under CNY 100,000, integrating liquid cooling is not as simple as "just installing it."
Wang Xingxing's remarks at the October Industrial Expo already clarified the underlying logic: The onboard computing power deployable in the robot body is relatively limited, constrained by "body structure, battery, size, and heat dissipation." Adding another high-performance chip to the body increases power consumption and heat generation, requiring larger cooling modules. These cooling modules occupy space and add weight, which in turn reduces battery life. Computing power, battery, and thermal management compete for limited space and power budgets within the same body—a nearly zero-sum game. This climbing competition merely brought this physical reality to the forefront.
02、Why Did the A2 and E04 Finish?
In the same competition, the finishers provided a different reference point.
While the Unitree humanoid robot stumbled during the climb, its A2 quadruped robot won the legged racing category with a final time of 43 minutes and 41 seconds (actual time41 minutes and 41 seconds, including a 2-minute penalty). It operated entirely in autonomous mode, perceiving and deciding in real-time, moving stably and quickly across irregular stone steps.
The A2 weighs approximately 42 kg, with peak joint output torque of about 180 N·m, and also uses localized air cooling. However, there is a fundamental difference in thermal management between quadrupedal and bipedal humanoid configurations: Quadruped robots have more uniform joint load distribution and relatively ample internal space, resulting in significantly lower sustained thermal density per joint compared to the repeated bursts required by humanoid lower limbs when climbing stairs. The fact that the A2 could handle this with air cooling does not mean the same thermal management logic can be transferred to the G1.
The winner of the humanoid racing category, Team Zhi Shen Ling Yue's E04, completed the race in 27 minutes and 26 seconds. The main race segment ran from Hongmen Square to Doumu Palace, while the challenge segment from Doumu Palace to Hutian Pavilion took 55 minutes and 34 seconds.
Public information did not disclose details of the E04's thermal solution. However, we can refer to another confirmed case: Honor's "Lightning," which won the Yizhuang Half Marathon this year, featured a self-developed liquid cooling system using a combined liquid and air cooling approach, with a heat exchange flow rate exceeding 4 liters/minute. This relied on relatively ample internal space and continuous active circulation.
Liquid cooling in humanoid robots will likely first be implemented in industrial-grade, high-load, long-duty-cycle platforms, rather than general-purpose models pursuing low cost and lightweight design.
03、Mount Tai Tests Transient Loads, Not Peak Values
If the issue were simply "climbing stairs strains joints," the solution would be straightforward—just make the cooling module bigger. The true cruelty of Mount Tai lies in the severe transient loads it imposes, whereas most humanoid robots' thermal management systems are designed for steady-state conditions.
When running on flat ground, joint motor output is relatively predictable, allowing the cooling system to operate on a stable thermal cycle. On Mount Tai's stone steps, foothold depth varies, surfaces are mossy and slippery, and the robot's center of gravity adjustment changes drastically with every step. Joint loads fluctuate repeatedly between peaks and valleys. While the target ratio of peak torque to continuous torque typically needs to exceed 3:1, systems with inadequate thermal design often achieve only about 2:1—resulting in insufficient peak torque or overheating during continuous output.
The G1's knee joint peak torque of 120 N·m looks impressive in an exhibition hall, but during continuous ascent, what truly determines whether it can proceed is the thermal boundary of continuous torque, not the peak value.
This explains why "switching to a larger liquid cooling pump" may not solve the Mount Tai challenge. Thermal management strategies must upgrade from "passive heat rejection" to "predictive heat generation based on tasks," coordinating with motion control algorithms to dynamically adjust power output according to temperature rise and load status. Simulation environments can precisely model cooling curves and ground friction coefficients, but real Mount Tai steps are wet, narrow, and mossy—physical parameters outside the training distribution. No matter how smart the AI motion control algorithm is, when the motor overheats and throttles, the execution layer fails.
04、The Other Side of the Stretcher
In the first half of this year, China's humanoid robot product varieties exceeded 400, with production surpassing 40,000 units; annual production is expected to break 100,000 units. Figure's production line has accelerated from one unit per day to one unit per hour, and AgiBot's 10,000th robot rolled off the assembly line. Capital market reactions were restrained: On the day the G1 was carried down the mountain, Unitree's stock opened low but closed up, gaining over 2%.
However, another figure from the CAICT report deserves attention: Global humanoid robot shipments exceeded 22,000 units in the first half of the year, with over 60% going to entertainment, commercial performances, and educational data collection. Only about 13% were deployed in actual intelligent manufacturing scenarios on production lines, and a significant proportion of these devices had average monthly operating days of less than 10.
Humanoid robots perform well in controlled environments, but once they leave structured spaces, reliability degrades non-linearly.
The G1 remains an excellent humanoid robot. But bridging the gap from "performing cool actions" to "operating continuously for eight hours in unfamiliar environments without systematic failures in thermal management, endurance, or gait" requires a comprehensive, system-level advancement in thermal management, power management, and reliability engineering.
CCTV positioned this competition as "real-world practical training," aiming to accumulate real data for mountainous intelligent equipment. That stretcher on Mount Tai is the most honest preview of this industry's current state.
This article is compiled based on publicly available online information, for reference only, and does not constitute investment advice.