The whistle blew, the scoreboard stopped at 9:0, and the accelerated evolution humanoid robot T1 won the championship trophy of the adult group final of the RoboCup German Open. Cheng Hao realized the championship dream that started with the Tsinghua Vulcan Team.
RoboCup is known as the "Robot World Cup". The Tsinghua Vulcan Team was founded by Zhao Mingguo and has participated in the competition for many years. The members come from his robotics laboratory. Cheng Hao, who graduated from the Department of Automation at Tsinghua University, was the third captain. His experience of leading the team to win the third place in the humanoid group Teen-Size during his student period was both the starting point of his robot dream and a lingering regret.
His career trajectory after graduation seemed to deviate from his original intention: Cheng Hao founded the "Chao Xi Calendar" and became the vice president of Feishu products after being acquired by ByteDance. However, the technological breakthrough of ChatGPT3.0 in 2023 allowed him to keenly capture the key turning point. The combination of the emergence ability of large models and robot motion control is reconstructing the underlying logic of embodied intelligence. Cheng Hao realized that the opportunities created by large models for humanoid robots are different from any previous moment in history. "General AI has emerged and will develop into multimodal, with perception and influence on the physical world, requiring new hardware carriers."
This sense of smell made Cheng Hao founded Accelerated Evolution in August 2023, recruiting many former Vulcan team members as R&D personnel, and invited his teacher Zhao Mingguo to join as chief scientist. About a year later, Accelerated Evolution's first humanoid robot product T1 was "born", confirming his forward-looking judgment.
As soon as T1 was "born", it "returned" to the football field familiar to Cheng Hao almost immediately. This entrepreneurial youth, who was born and raised in Beijing, is determined to start from the football scene and gradually realize the intelligent generalization of humanoid robots.
"The reinforcement learning of robots is similar to that of children learning to walk, from crawling to standing up and walking, from staggering to walking steadily, and then to climbing slopes, steps, and even jumping. It's just that in the virtual world, training several robots for a few hours is equivalent to 10 years for humans." Cheng Hao told reporters.
T1 won the RoboCup World Cup for the second time
In March, in the adult final of the RoboCup German Open, the accelerated T1 faced the host German Sweaty as the blue team. On the field, a T1 in the front seemed to "predict" the opponent's blocking route, bypassed the red team that was a head taller than itself, touched the ball first, and scored a goal. The score of this duel was finally locked at 9:0, and the Tsinghua Vulcan team won the championship with a zero seal. Cheng Hao refreshed his own participation record.
In July 2024, he returned to the RoboCup field again in the name of the Tsinghua Vulcan team. Compared with his student days, Cheng Hao is no longer just a robot developer. From the laboratory to the arena, he has completed a key transition in cognitive paradigm. At this time, T1 could not persist in "playing" the entire game. Many robots "lamed" and fell because of joint heat. When colliding with other robots, there was even a disconnection between the main control and the joints, and the entire robot was powered off and shut down. "If it breaks, you have to replace it immediately." Cheng Hao prepared 10 robots for a 2V2 match, but eventually lost to the champion American team.
Cheng Hao was not discouraged by not winning the first game. "If you are not tall enough, you can make up for it with your skills. Dominating the court is not a dream!" After the game, this sentence was written into the summary article of the WeChat public account of Accelerated Evolution as T1's motto.
"The plastic chest shell and head shell of the robot will crack after being bumped." This gave Cheng Hao a headache. Optimizing materials, continuing experiments, no less than dozens of trial and error, even using a 10-20 kg ball, falling from a height of two meters, Cheng Hao finally found "tough" and fall-resistant materials in a ruin. Cheng Hao mixed them into motors, reducer gears, bearings, brackets and other components, and T1 had a new "skin".
Friction and collision are inevitable in court competitions. In order to allow the main control to work stably and not lose power under the high vibration state after wrestling, Cheng Hao optimized the design of embedded hardware and specially developed vibration and drop test equipment. This device can hit at a frequency of two peak forces per second for 20 hours. "We will select a few joints from each batch for testing. If they can survive 20 hours of impact without damage, they will meet the requirements."
How can a robot get up quickly after being hit? To this end, Cheng Hao continuously iterates the algorithm for getting up. It used to take T1 10 seconds to get up, but now it can get up in one second, and the success rate has also approached 100% from 50%.
In an experiment, 6 T1s played 3V3 against each other. After 20 consecutive games, most of the joints were not damaged and no "substitutes" were needed to save the game. After the experiment, T1 was full of confidence and "returned" to the RoboCup arena again, becoming "tough" and resistant to falling, and "flexible" to get up, winning the championship cup in one fell swoop.
On the other hand, most of the contestants (robots) still couldn't stand the collision. For this reason, the event temporarily adjusted the rules from allowing certain collisions to prohibiting collisions.
The first stage of the Zhongguancun Forum was amazing
T1 was officially unveiled in August 2024. Cheng Hao said that with the support of the Beijing Municipal Government, this smart device carrying cutting-edge technology has quickly become a "city technology business card", frequently attending international events as a "facade" to show its prowess, and will also be stationed in iconic venues such as the Water Cube to achieve technical iteration and upgrades through high-standard robot competitions.
In October 2024, T1 began mass production and shipment, and completed the delivery of the 100th unit in early March this year, marking that T1 has officially entered the fast lane of industrialization.
The Zhongguancun Forum in March this year became a stage for displaying its technical strength. T1 "participated" in the Zhongguancun Forum for the first time, accelerating its evolution. More than a dozen machines were dispatched to appear together. Some of them appeared on the opening stage, performing push-ups and standing up on the spot, while others stood at the entrance of the main forum to welcome guests, acted as hosts at sub-forums, or played football in the exhibition hall, and perfectly presented the collective wisdom of artificial intelligence through cluster collaborative operations.
From a technical perspective, T1 uses a reinforcement learning motion control algorithm on the robot's "legs". Tens of thousands of robots are trained in a simulated data environment at the same time, "walking" through various environments such as steps, slopes, smooth surfaces, and rough roads. Cheng Hao told reporters, "The reinforcement learning of robots is similar to that of children learning to walk, from crawling to standing up and walking, from staggering to walking steadily, and then to climbing slopes, steps, and even jumping. It's just that in the virtual world, several robots training for a few hours is equivalent to 10 years for humans."
Cheng Hao pointed out that robot learning also has a reward and punishment mechanism, "but the model is more complicated." Only the result model with good training performance has the opportunity to be deployed on the real machine, and a series of algorithms are used to reduce the gap from simulation to real environment. "Our current motion control algorithm not only controls the legs, but also hopes to control the joints of the whole body to move together, such as making a powerful shot. This requires the introduction of imitation learning algorithms, allowing robots to imitate stickmen in a simulated environment, and then consolidate through reinforcement learning."
The embodied intelligence large model still has shortcomings but belongs to the "future" of the industry
Cheng Hao loves to play football, and his robot also plays football.
"Robots playing football is a highly efficient experimental scenario." Cheng Hao believes that robots playing football require visual ability. In addition to identifying their own field position, they also need to identify objects such as the ball and the goal. At the same time, they also need decision-making ability to judge whether they are forwards, midfielders, or defenders, "what they do will be different." In addition, they must constantly adjust their decisions and cooperate with each other as the opponent's robots move. Robots playing football also emphasize motion control, and require high athletic ability such as walking, turning, shooting, and getting up. "Just like football players, their qualities are relatively comprehensive."
In Cheng Hao's eyes, football and life scenarios have similar requirements for some perception and decision-making underlying algorithms, such as object recognition, tracking, positioning, and obstacle avoidance in decision-making. He believes that "starting from football scenarios does not require too much worry about safety issues, and it is relatively economical."
"Through robots playing football, we are constantly researching new motion control, perception, and decision-making algorithms." Although Cheng Hao believes that the current embodied intelligence large model still has shortcomings, it belongs to the future direction of the industry.
"I think it is easier to start a business in Beijing and expand it. The cost of manpower and site may be more expensive, and the team needs to make better income and financing." Cheng Hao said. In 2024, Accelerated Evolution received a Pre-A round of financing jointly funded by Zhongguancun Science City and other institutions. The number of Beijing base model companies, talent density, and scene support also gave him confidence in the future.
He believes that Musk's statement that humanoid robots will reach 10 billion units, and there will be multiple models at the same time, "factories, companies, and families have them", but this may take 10 or 20 years. Cheng Hao remembers that the first computer in the community when he was a child was worth 10,000 yuan, "at that time, a house in the community was also 10,000 yuan."
But computers, which were once equivalent to buildings, are no longer out of reach. The sound of keyboards can be heard one after another in office buildings, and even the convenience store downstairs can use computers to check out. In the future, humanoid robots may also be able to stay in the street window like "Clara and the Sun", waiting for Qiao Xi to take them home.