Chinese Humanoid Robots Play Tennis and Drive Karts in New Demonstrations

In the tennis event, the robot was shown autonomously moving, judging incoming balls and striking them with forehand and backhand shots, serves, returns and volleys. It engaged in multiple continuous rallies against human players and collaborated with a human teammate in doubles, adjusting its tactics in real time. Even after falling during high-speed play, it got up and continued. The company said this was the first time a humanoid robot had completed a fully autonomous tennis match in a large-scale live broadcast. The performance was powered by its self-developed embodied intelligence model, the Galaxy Xinge (AstraBrain), which integrates high-level decision-making and low-level motor control into one system, according to the report.

Symbiotic Intelligence's video showed a different but equally challenging skill: a dual-legged humanoid robot seated in a go-kart, driving at full speed through turns and navigating obstacles in a single unedited shot. The team said this demonstrated whole-body coordination of eyes, hands and feet under high-speed motion and changing postures, as well as precise force control during multi-contact balance. The company, founded just two months ago, is developing an end-to-end “brain” for dual-legged humanoid robots. Its founder, Ding Pengxiang, born in 1996, said the goal is to make robots behave more like humans rather than perform only mechanical services.

Symbiotic's technical approach is notable for rejecting the layered architecture common in current systems. In layered models, a high-level “brain” decides tasks and a low-level controller handles motion. Ding and his team argue this creates information bottlenecks and makes true scaling impossible. They are pursuing a pure end-to-end model that directly outputs joint targets, while using a dual-domain optimization mechanism to combine task accuracy with stability learned from low-level controllers. The team, all PhD students, has published over 40 top-conference papers and built one of China's first embodied foundation models, according to QbitAI.

Both demonstrations point to rapid progress in making humanoid robots operate in dynamic, unstructured environments. The tennis match represents a milestone in physical AI, showing that robots can perceive, decide and act in real-time against human opponents. The kart-driving demo highlights the challenge of controlling a bipedal body that must maintain balance while manipulating a vehicle. The startups are betting that end-to-end models trained on large-scale data will eventually outperform modular systems, though the industry has not yet converged on a final architecture.