At the second World Humanoid Robot Games in Beijing, Team Tiangong's Omni humanoid robot discovered a running technique it was never programmed to use, teaching itself through trial and error in simulated environments. The robot realized that the typical human arm-swinging motion wasted battery power and caused excessive joint friction, so it invented a tucked-arm "Shy run" instead, compensating with rapid waist-twisting movements to maintain balance. This breakthrough represents what experts are calling a "Move 37" moment for physical artificial intelligence (AI), comparable to when Google's AlphaGo made an unprecedented move in 2016 that no human Go player would have considered.

What Makes This Robot's Discovery So Significant?

The Omni robot's innovation matters because it demonstrates that AI can do more than follow human instructions in the physical world. Just as AlphaGo discovered strategic moves outside human intuition in the digital realm, this humanoid discovered an optimization strategy that humans hadn't conceived of. The robot coordinated its joints, torso, and center of gravity to solve a real engineering problem: how to run efficiently while managing energy consumption, heat generation, and mechanical stability simultaneously.

This breakthrough was only possible because of the robot's advanced hardware. The mechanical design, balance systems, joint performance through actuators, and heat management capabilities had to be sophisticated enough to support the robot's self-discovered running style. Without these foundational technologies working in concert, the AI's optimization strategy would have been impossible to execute.

How Does Reinforcement Learning Enable Robot Innovation?

The Omni robot used reinforcement learning, a machine learning technique where AI systems learn through trial and error in simulated environments before applying what they've learned to the physical world. The robot experimented with thousands of different running motions in simulation, testing which approaches conserved energy, reduced heat, and maintained stability. Once it identified the most efficient pattern, it transferred that knowledge to its physical body.

This approach differs fundamentally from traditional robot programming, where engineers manually code every movement. Instead, the AI discovers optimal solutions based on the constraints it faces. The robot essentially asked itself: "Given my physical limitations and energy budget, what's the best way to move?" and found an answer that surprised even its creators.

What Are the Broader Implications for Physical AI Development?

The World Humanoid Robot Games revealed the scale of progress in the physical AI ecosystem. The second annual competition brought together 2,056 robots from 666 teams across 16 countries and regions, more than quadrupling the 500 robots from 280 teams that competed in the inaugural event one year prior. The games expanded from 26 events to 51, including 30 competitive events and 21 scenario-based challenges spanning industrial assembly, hospitality, household services, firefighting, and rescue operations.

Each task demonstrated the complexity of coordinating multiple systems simultaneously. A humanoid robot requires AI computing and control software to make decisions, sensors and machine vision to interpret its surroundings, and motors, actuators, reducers, batteries, and mechanical systems to translate that intelligence into movement. The games made this entire ecosystem visible by showcasing robots performing diverse tasks.

Which Real-World Applications Are Already Showing Results?

Beyond the Omni's running breakthrough, other robots at the games demonstrated practical capabilities in commercial scenarios. In the power-tool assembly challenge, LinkerBot won gold by autonomously installing 18 screws in five minutes with no human intervention. In the restaurant-service event, a robot from GalBot completed a full autonomous service sequence, using the company's AstraBrain embodied AI model to make decisions based on voice commands. The robot located and microwaved food, prepared drinks, and delivered a fully correct order to a designated table in 6 minutes and 55 seconds with precise force control and no errors.

These demonstrations show that physical AI is moving beyond laboratory settings into real-world applications where robots must handle unexpected situations and make autonomous decisions. The Smart Charging Service challenge, for example, required robots to unplug charging connectors, charge three vehicles, and return each connector to its original position within 30 minutes, tasks that demand both precision and adaptability.

How to Understand the Physical AI Technology Stack

• AI Computing and Control Software: The decision-making layer that processes information and determines what actions the robot should take based on its environment and objectives.
• Sensors and Machine Vision: The perception systems that allow robots to see, measure, and interpret their surroundings in real time, enabling autonomous navigation and task execution.
• Motors, Actuators, and Mechanical Systems: The physical components that translate AI decisions into actual movement, including joints, batteries, and structural elements that must work together seamlessly.

What Do Industry Analysts Predict for Humanoid Robot Adoption?

The momentum demonstrated at the World Humanoid Robot Games aligns with bullish forecasts from major financial institutions. Goldman Sachs recently revised its humanoid shipment estimates upward more than fourfold, raising its base case projection to 6.5 million units by 2035, up from a previous estimate of 1.4 million units. This dramatic increase reflects growing confidence that humanoid robots will integrate into workplaces and services at scale within the next decade.

The opportunity extends beyond individual robot manufacturers. Competing robots may differ in design and developer, but they generally rely on the same foundational technologies and components. As capabilities improve and commercial use cases broaden, value creation is occurring across the entire physical AI stack, from component suppliers to software developers to the robot makers themselves.

The Omni robot's self-discovered running technique represents a watershed moment for the field. It proves that physical AI systems can optimize real-world performance in ways humans might never conceive, solving engineering problems through machine learning rather than manual programming. As the World Humanoid Robot Games demonstrated, this capability is no longer theoretical; it's being deployed in competitions and commercial applications today.