During the opening ceremony of the Second World Humanoid Robot Games in Beijing on August 23, 2026, humanoid robots built by Beijing Galbot Co., Ltd. completed more than 100 consecutive autonomous tennis rallies against human athletes, marking a significant milestone in embodied artificial intelligence.
From Board Games to the Baseline: The AstraTennis Milestone
Ten years after DeepMind’s Go program defeated Lee Sedol in a fully observed, rules-bounded digital environment, artificial intelligence is stepping into the physical world. Galbot framed its live exhibition as the “AstraTennis” moment. Unlike scripted walking routines or teleoperated demos that dominate the robotics circuit, this event required the company’s humanoid hardware to handle a continuous, physical, and partially observed domain.
Tennis forces a robot to track a fast-moving target, plan a full-body response, coordinate locomotion and a racket swing, and adjust to a human opponent inside a fraction of a second per shot. According to the company’s announcement, the match featured Galbot robots autonomously tracking high-speed tennis balls, positioning themselves on the court, and executing serves, forehands, backhands, returns, baseline rallies, net play, and recovery shots.
The core breakthrough claimed by the company is true autonomy. The hardware was not relying on pre-programmed trajectories or remote human operators. Instead, the onboard systems perceived the game state, selected shots, and adapted strategies in real time.
Hardware Resilience and Real-Time Doubles Play
Sustaining 100 consecutive exchanges is a metric of stability. A single successful return proves very little about a perception and control stack. Keeping a rally alive past the century mark implies that the ball-tracking pipeline, court positioning algorithms, and swing timing held up under live conditions.
The integration of a doubles format pushed the operational complexity further. Partnering with human tennis champions, the robots shared the court, covered their designated halves, and adjusted their positioning based on both their partner’s movements and the opponents’ shots. This mimics the unstructured cooperation required in real-world industrial and retail workspaces.
Physical stability during fast exchanges remains a major hurdle in humanoid engineering. According to the company, the robots lost balance during high-speed exchanges but executed unassisted recovery moves, getting back on their feet to continue competing without interruption.
Behind the tennis exhibition is a commercial platform. Galbot, formally known as Beijing Galbot Co., Ltd., has primarily positioned its hardware—such as the wheeled dual-arm G1 humanoid—for retail, manufacturing, and pharmacy work.
Evaluating the Physical AI Stack
The transition from virtual LLM training to embodied robotics requires sub-millisecond sensor fusion. Tracking a tennis ball travelling at high speeds demands rapid frame-rate processing and low-latency motor actuation. While board games like Go operate on discrete turns, tennis is continuous and unforgiving of latency spikes in the control loop.
Galbot’s demonstration highlights the gap between static automation and adaptive physical intelligence.
As embodied AI transitions from controlled lab environments to dynamic real-world arenas, exhibitions like AstraTennis set a new baseline for what humanoid platforms must achieve in terms of physical perception, balance recovery, and real-time decision-making.
