World Robot Mount Tai Competition Reveals Technical Challenges for Domestic Robots

Ask AI · Mount Tai competition exposes shortfalls in heat dissipation, how to iterate and optimize the energy system?

【Global Times Report, Reporter Chen Zishuai】Editor's Note: Recently, the first World Robot Climbing Mount Tai Competition was held in Tai'an, Shandong. Humanoid robots and legged robots successively challenged steep steps, not only competing in mountain climbing speed but also simulating Mount Tai porters carrying and transporting materials, testing motion control, environmental perception, and load capacity in a real mountain environment. Moving from the laboratory to complex outdoor scenes, robots climbing Mount Tai demonstrated the progress of domestic robot technology and also exposed a series of issues worthy of attention: What technical hurdles must be overcome to climb Mount Tai? In which fields can the capabilities demonstrated by robots in the competition be applied? In what directions will robots evolve in the future? Around these questions, Global Times reporters interviewed several experts for interpretation.

Humanoid robot climbing Mount Tai. (Visual China)

Core difficulties: Must pass three hurdles first

In this competition, the humanoid robot racing official stage covers a total distance of about 1.6 kilometers and over 600 steps. The bipedal robot racing covers the entire distance from Taishan Hongmen Square to Zhongtianmen, about 3.9 kilometers and over 3100 steps. What capabilities of the robots will such a scene test? “Robots climbing Taishan, the real challenge is not climbing steps, but achieving continuous, stable, and autonomous movement in complex, unstructured real-world environments.” Shenzhen Institute of Artificial Intelligence and Robotics Embodied Intelligence Director of the Center, Liu Shaoshan, told reporters from Global Times that there are three core technical difficulties behind this. First is environmental perception, where the robot needs to identify steps of varying heights, ground friction conditions, and footing positions in real-time. Second is dynamic control, where the robot needs to constantly adjust its center of gravity, joint force, and gait during movement to avoid losing balance. Third is system reliability, where continuous climbing of thousands of steps places high requirements on the real-time performance of the motor, energy, heat dissipation, and computing system.

“This competition reflects the obvious progress of our robots in complex terrain motion control and long-distance continuous motion. Especially the quadruped robots, which have already demonstrated strong mountain passage capabilities; humanoid robots are also developing from standardized gaits in laboratories toward dynamic motion in complex environments.” said Liu Shaoshan.

It is worth noting that the competition also set up a load-bearing race for quadruped robots, which simulates the work scene of a Taishan mountain porter, competing in the material transport capability of robots under a load-bearing state; the entire competition section is about 3.3 kilometers long with over 2,100 steps. Zhong Xinlong, Director of the Artificial Intelligence Laboratory of the Future Industry Research Center of Saidi Research Institute, told Global Times reporters that the load will increase the force on the joints and affect the center of gravity and balance as the position of the cargo changes. The back of a quadruped robot can be equipped with a relatively stable cargo platform, and multiple support points are also conducive to sharing the weight, therefore it is more suitable for prioritizing the testing of mountain transportation capabilities.

Why not let humanoid robots carry loads while climbing Mount Tai? Zhong Xinlong stated that when humanoid robots carry loads on two legs, the upper limb movements will change the center of gravity of the entire machine, posing higher requirements for sole friction, continuous joint torque, and fall protection. “This competition reminds us that the robot's form should be determined by the specific task. For mountain transportation and slope inspection, quadruped robots often have more advantages at the current stage. For opening doors, picking up objects, and operating tools, especially in facilities built according to human height and movement habits, the hands and upper bodies of humanoid robots have more room to play. The future industrial landscape is very likely to have different forms performing their respective duties.” he said.

Future applications: Operating in complex environments

What technical skills demonstrated by the robot climbing Mount Tai can be applied to which fields and scenarios in the future? In Liu Shaoshan's view, the most direct application directions are inspection, transportation, and emergency operations in complex environments. “For example, mountain inspection, power facility maintenance, material transportation at disaster sites, as well as operations in dangerous environments such as mines and tunnels. These scenarios have practical requirements for navigation in complex terrain, and there is also commercial value in reducing labor intensity and safety risks for personnel.”

A reporter from the Global Times noted that the industry has already begun exploring relevant applications. For example, the Taishan scenic area previously conducted tests for robot dog waste transport. In October, a Swiss industrial robot company announced it had begun delivering four-legged inspection robots for hazardous environments to customers in oil, gas, and chemicals. Liu Shaoshan said that four-legged robots already have an application foundation in some limited scenarios, and there is hope to further expand commercial deployment in the next one to three years.

In the field of humanoid robots, applications may start more from factory material organization, warehouse picking, and fixed tasks requiring two-hand operation, and then gradually move toward positions with more complex environments. Currently, domestic robot companies have already conducted joint trials with automobile factories and others. Professor An Bo, Chair Professor at Nanyang Technological University in Singapore, Dean of the Interdisciplinary Research Institute of Artificial Intelligence, and Head of the Department of Artificial Intelligence, told a Global Times reporter that four-legged robots are already being used in relatively orderly scenarios such as substation and industrial park inspections. Humanoid robots entering complex outdoor environments and being used on a large scale still have to wait a few years.

Liu Shaoshan mentioned that the current largest technical goal is for robots to be able to autonomously complete valuable work, and this competition is a milestone in reaching this goal. “This goal involves not only motion capabilities but also environment understanding, task planning, dexterous manipulation, continuous learning, as well as the reliability and economic feasibility of the entire system.”

An Bo said that it is not difficult to run a competition once, but actual applications require it to work hundreds of times every day in wind and rain without issues, which are two different matters. Another is endurance and load; battery energy density is an old problem, and you cannot walk very far while carrying things. There is also one point that is often ignored: climbing the mountain is just arriving at the scene; after arriving, whether you can understand what needs to be done and whether you can actually do the work is what the users truly care about.

A reporter from the Global Times noticed that the robot did not have a smooth journey during the mountain climbing process, with situations such as staggering, pausing, and manual pulling occurring, and even a robot withdrawing from the competition due to “heatstroke” from overheating.

Zhong Xinlong pointed out that the competition also exposed the shortfalls of the robot, as the robot's mobility, energy supply, and safety control have not yet been fully coordinated. He believes that to fill these shortfalls, first, the continuous output capability must be incorporated into the joint and overall machine design, improve the heat dissipation efficiency of the motor, reducer, and drive system, and allow the robot to actively adjust its speed and gait based on temperature, remaining battery power, and task distance. Second, repeated tests must be conducted under conditions such as slippery steps, gravel roads, long slopes, and heavy loads, focusing on improving the ability to re-foot, safely stop, and autonomously recover after perception errors. Finally, the testing standards should be changed from single-pass completion to entire-shift operation, assessing continuous working time, effective transport volume, and so on.

Key Turning Point: From “Active” to “Capable”

“Can run, can jump, can climb mountains, it indicates that the robot possesses better body control capabilities; whether it can understand tasks, adapt to unknown environments, and autonomously solve problems is the higher standard for measuring embodied intelligence.” Liu Shaoshan said, there are three core challenges in the next stage. The first is Generalization Capability.Robots cannot only perform excellently in trained environments; they also need to adapt to new environments, new objects, and new tasks. Second is autonomous learning capability. Robots need to gradually acquire the ability to use historical experience and current environmental information to solve new problems, reducing the dependence on large-scale manual data collection and repetitive training. Third is computing system capability. Future robots need to simultaneously run different types of computing tasks such as visual perception, environmental understanding, world model, task planning, and real-time motion control. These tasks have different requirements for computing power, memory, latency, and security, and existing computing architectures still have significant room for optimization. Therefore, the development of embodied intelligence cannot rely solely on more powerful artificial intelligence models; it also requires collaborative innovation between models, computing chips, control systems, and the robot body.

An Bo also believes that robots are currently at a critical turning point, moving from being "active" to being "capable," and from "single skills" to "general intelligence." In recent years, with the help of reinforcement learning, large-scale simulation training, and "simulation-to-reality" transfer technology, the movement capabilities of robots have advanced rapidly; running, jumping, overcoming obstacles, and climbing stairs can all be achieved now, and this Taishan competition is a concentrated display of the achievements of this stage. It can be said that at the "cerebellum" level, that is, in motion control and body balance, we have already achieved considerable breakthroughs.

“But the real short board is at the ‘brain’ level. Today, most robots still perform excellently in relatively clear tasks, but once the environment changes or the task is open, it is difficult for them to understand intentions, make plans, and respond flexibly like humans. For example, if you ask it to move a box to the top of a mountain, it can do it, but if you say ‘see what is missing on the mountain and fill it in,’ it wouldn't quite understand.” An Bo told Global Times reporters, large language models and multimodal large models The emergence of these brings new possibilities to this problem; they allow robots to understand natural language instructions, comprehend scenes, and perform common-sense reasoning. Combining this type of “agent brain” with a robot body, which is embodied intelligence, is one of the most active directions of global research currently. However, how large models can reliably interface with the precise control of the physical world, how to obtain enough real interaction data, and how to ensure decision safety all need to continue to be focused on.

As for the future development of robots, An Bo summarized three directions. First, moving from individual intelligence toward collective synergy. In real-world scenarios, there are very few cases where a single robot completes a task alone; it is often multiple robots of different forms, such as quadrupedal, wheeled, and drones, plus humans, completing the task together. Second, moving from the laboratory toward the open real world. Future robots need to possess the ability for continuous learning, constantly accumulating experience and self-improving during deployment, rather than being re-developed every time a scene changes. Third, moving from technological breakthroughs toward safety, trustworthiness, and human-robot integration. How to make robot behaviors predictable and capable of being supervised and corrected by humans, and how to prevent risks in environments where multiple autonomous systems coexist, are problems that the technical community, the industrial community, and regulatory departments need to face together.