The World’s First Humanoid Robot Combat Tournament Concludes: Fierce Exchanges, but Cognitive Systems Still Lag Behind

At the CMG World Robot Competition · Series, Unitree Technology participated as a partner. All robots on stage were Unitree G1 humanoid robots, while the embedded algorithms were developed independently by four competing teams and controlled on-site via handheld controllers. After several rounds, the Black Team “AI Strategist” secured the championship title.

According to Unitree, the combat actions used in the event were sourced from professional fighters to train the AI models. Despite current limitations in dynamic control and real-time perception, the robots demonstrated notable balance and coordination in human-machine interaction.

Industry experts told KeChuangBan Daily that while the underlying algorithms have matured, the absence of a spatial intelligence large model—acting as the robot's "brain"—remains a critical gap. Future development may see more solution providers emerge to create specialized systems on top of public-version robots.

During combat, the robots adjusted footwork to maintain balance and could get back up quickly after a fall. One audience member commented, “Though the fighting wasn’t spectacular, I was impressed by how agile it was when getting up. The support strength was maxed out.”

Tian Feng, Dean of Think Slow Think Fast Institute and Co-Director of the Computational Law and AI Ethics Research Center at Shanghai Jiao Tong University, noted: “A robot that can rise within five seconds and maintain balance when unstable is impressive. Dynamic balance in bipedal robots is far more difficult than in quadrupeds. A quadruped only needs to keep its center of gravity within a rectangular base, while bipeds must maintain it between two feet.”

Unitree explained that the G1 robot relies on reinforcement learning in simulated environments and joint-level perception to maintain balance. The robot trains with vast balance data and simulations, optimizing its balance strategies. High-precision joint sensors track posture and force distribution in real time, providing critical input for balance control.

The IMU (Inertial Measurement Unit) plays a crucial role in maintaining equilibrium, offering real-time posture and acceleration feedback during both actual and simulated operation.

Notably, the G1’s arms were custom-designed for the event, featuring seven degrees of freedom—two more than the typical five—to allow more versatile techniques, such as hooks, swings, and uppercuts. More flexibility and assault angles are made possible by these improvements.

According to Unitree, training the G1 for combat presents three main challenges: high torque and rapid response demand a powerful drive system; real-time sensor data processing requires precision control algorithms; and the mechanical structure must withstand significant impact, posing strict requirements on joint and skeletal strength.

G1 robots support multiple control modes, including voice, handheld, and motion-sensing. The tournament used primarily handheld controllers for live operator input.

Voice control has substantial latency, which limits responsiveness, according to Unitree. Handheld controllers offer intuitive and precise control, making the system more accessible to participants. Motion-sensing control provides immersive potential; Unitree has already developed a corresponding system that may be used in future competitions.

Liu Tai, Deputy Chief Engineer at the CAICT’s Tyre Systems Lab, explained that handheld control of humanoid robots differs from standard toys. A complete algorithmic system underpins control, including large models and motion control logic.

“Conventional control methods couldn’t ensure stability. Reinforcement learning allows the robot to discover balance strategies independently. The competition was inspiring—robots maintained coordination and stability under pressure. It marked clear progress,” Liu stated.

Tian Feng added, “In the absence of AGI, the industry is exploring human-machine collaboration. This tournament has shown promising outcomes in that direction.”

Human-controlled robots still encountered errors. For example, mistimed attacks led to falls and knockouts, and robots became entangled in the ropes while attempting to reposition—requiring human intervention. These incidents highlight current limitations in dynamic control and real-time perception.

Tian Feng posed the question: “Should robots replicate human anatomy? Humans have no eyes on their backs—should robots have rear cameras to improve perception? Humans have five fingers, but some industrial applications might benefit from six. Humanoid robots may be transitional; future robots could surpass human-like forms.”

An industry attendee commented that current underlying algorithms already excel in areas like stability, vision, and agility, but a spatial intelligence large model is still missing.

Tian Feng observed that compared with the U.S., Chinese investment is more focused on robot hardware, while American firms are more engaged in developing robot “brains.” “Most domestic companies develop the body. Some work on the 'cerebellum'—motion control. Only around 1 in 20 are developing the brain. But some have started already.”

He further noted that while industry focus has largely been on complete machines and components like motors, foundational software is equally vital. Secondary development is essential for adapting robots to combat, manufacturing, and logistics. The future may see more companies developing intelligent solutions on top of standard models—integrating multi-modal models and RAG-based knowledge systems.