On the evening of August 22, the second World Humanoid Robot Games opened at the National Speed Skating Oval. In a live broadcast, a dual-legged tennis-playing humanoid robot from Galaxy General stood at the baseline, facing off against tennis star Zheng Jie. From serves and sprints to forehand and backhand switches, baseline rallies, net volleys, doubles coordination, and even getting up after a fall, the robot completed a full-scale match in a real-world setting. That moment has been dubbed “AstraTennis.” A decade after AlphaGo’s victory over Lee Sedol on the Go board, this humanoid robot caught a professional player’s return on the court, marking what experts call a shift from digital intelligence to physical AI.

No Remote Control, No Pre-Set Paths: A Complete Match

A tennis court is far from a controlled laboratory environment. With ball speeds exceeding 50 km/h, the robot’s binocular vision system must lock onto the ball’s trajectory within 0.1 seconds while adjusting its posture mid-run, judging the landing point, controlling wrist force, and choosing a hitting strategy. Throughout the exhibition, the Galaxy General robot displayed the full range of tennis skills—serving, forehands, backhands, returns, baseline control, and decisive net play—engaging in extended rallies with its human opponent. It handled different speeds and spins with ease, and in doubles, it coordinated effectively with its partner, using cross-court positioning and tactical communication to counter the opposition, demonstrating genuine teamwork.

The robot showcased nearly every ability of a real tennis player. It delivered precise serves, moved laterally with small adjusting steps, rotated its torso, and swung through shots with fluid, human-like mechanics. On difficult placements, it sprinted across the court for desperate saves, and even after falling, it rose independently to continue. During baseline exchanges, it controlled shot placement based on the opponent’s position, and when opportunities appeared, it moved forward to the net to alter its approach. Crucially, the robot remained unaffected by psychological pressure, focusing solely on winning each point regardless of the score. Behind all this, there was no remote control, no pre-programmed trajectories, and no human operator—every action depended on the robot’s real-time perception, decision-making, and execution.

Industry experts note that most bipedal robots struggle merely to walk steadily on flat ground, let alone perform high-speed running, sudden stops, and precise striking in uncontrolled settings, calling this a gap that cannot be closed in a single generation. In March of this year, Galaxy General first achieved fully autonomous tennis play, defeating human opponents with strategic left-right and front-back shots while maintaining balance and precise wrist control. By August, this live broadcast exhibition introduced far greater complexity and unpredictability—changing indoor lighting, electromagnetic interference from broadcasting equipment, and the acoustic pressure of tens of thousands of spectators—all variables absent from lab tests. Notably, when the robot fell during a difficult save, it quickly regained its footing using its autonomous balance algorithms and immediately resumed play without any external assistance, showing remarkable dynamic responsiveness.

Technology Roadmap Determines Capability Ceiling

Most humanoid robots on the market rely on a “brain + cerebellum” split architecture, where the brain handles environmental understanding and planning, and the cerebellum manages motor control, with each developed separately and then integrated. Galaxy General takes a different path. Its Galaxy Brain is the world’s first end-to-end embodied large model that unifies cognition, motor control, and neural regulation into a single system. This unified architecture means that understanding the environment, making decisions, controlling all joints, and adjusting posture in real time are handled by one model, eliminating any communication lag between “thinking” and “acting.”

Within Galaxy Brain, the system employs a World-Action Model architecture, integrating the action-generation capabilities of vision-language models with the predictive power of world models. The “brain” component is responsible for understanding and deciding: it assesses the ball’s direction, speed, and landing point, factors in the opponent’s position and match situation, and determines the type of shot, target placement, and next tactical move. In doubles and human-robot collaboration, it also reads teammate and opponent states to adjust tactics dynamically—essentially, it figures out how to win. The “cerebellum” component translates these decisions into physical actions. Tennis strikes occur in an extremely short window, requiring the robot to maintain full-body dynamic balance while sprinting, coordinating legs, torso, arms, and wrists with explosive force and accuracy. Through this drive, the robot’s movements become fluid and human-like rather than mechanical or fragmented.

The industry has long suffered from a one-model-per-machine, one-skill-per-scenario approach, where intelligence cannot transfer across platforms. Embodied AI also faces the common bottleneck of real-world data collection being too slow, expensive, and risky. Galaxy General addresses this with its Galaxy Studio data platform, which prioritizes synthetic simulation data supplemented by real-world data. The team generates billions of diverse scenarios in high-precision physics simulations, allowing virtual robots to engage in millions of multi-agent training matches in a virtual tennis world. Skills emerge in simulation and then transfer to the physical world. The company states that its tennis simulation training is equivalent to decades of continuous human practice. Experts note that while virtual-real fusion training is a popular direction in China, very few teams can achieve billion-scale data volumes with successful transfer to the physical world.

Validation Through Transfer: Scenarios as a Moat

One online observer commented after the exhibition: “The point isn’t that a robot can play tennis—it’s that it simultaneously runs all the technical modules needed for home services on a single machine.” Tennis demands perception, decision-making, motor control, and real-time strategic interaction, and combining these four dimensions is precisely the core bottleneck for embodied AI’s large-scale application. Experts suggest that once this capability is validated, the underlying skill base can be transferred to home services, industrial manufacturing, and healthcare.

When AlphaGo defeated a human champion on the Go board, AI achieved a landmark milestone in digital intelligence. A decade later, AstraTennis caught a professional player’s return on the tennis court. This time, the leap occurred in the physical world. For industry experts, AstraTennis’s significance goes far beyond a robot learning a sport. It marks the first time a self-developed Chinese embodied intelligent robot has demonstrated human-level perception, decision-making, action, and strategic play in a high-speed, open, real-time competitive environment. Galaxy General has carved out a genuine path of original innovation in embodied intelligence.