“Robot Just Beat Me at Badminton”: ANYmal-D Stuns Scientists With Inhuman Reflexes and Precision in Terrifying Display of Agility

The evolution of robotics has taken a significant leap forward with the development of the AI-powered ANYmal-D robot by ETH Zurich. This innovative robot is not only capable of playing badminton against human opponents but does so with a level of agility and precision that showcases the remarkable potential of modern robotics. By employing a reinforcement learning-based control system, this robot tracks, predicts, and returns shots with impressive accuracy. This breakthrough represents a crucial advancement in the integration of perception and movement in robotics, potentially setting the stage for future applications beyond sports.

Unified Motion Control

Playing badminton is a test of agility, precision, and coordination, requiring athletes to manage rapid footwork and precise arm movements. Reproducing these capabilities in legged robots has been a significant challenge due to limitations in current control systems and hardware. Human eyes, for instance, outperform commercial robot cameras in motion stabilization and focus, making the visual tracking of fast-moving shuttlecocks particularly complex. Previous research in athletic robotics has made strides in dynamic tasks such as flipping or running using reinforcement learning (RL), but these efforts often lacked integrated manipulation or relied on static environments.

“We Need It on the Front Lines Now”: Ukraine Eyes Czech Robot Drone Truck to Boost Troop Survival in Brutal Combat Zones

In an effort to overcome these challenges, researchers at ETH Zurich developed a unified RL-based control system for the quadrupedal robot ANYmal-D. This system autonomously integrates legged locomotion and racket swinging, allowing the robot to track, predict, and return the shuttlecock in real time. By using a perception-aware model trained in simulation, the system effectively accounts for motion-induced visual errors, narrowing the sim-to-real gap. The controller employs an asymmetric actor-critic framework, tracking multiple swing targets to learn continuous and responsive behavior. This innovative approach balances agility and visual accuracy, incorporating shuttlecock prediction, constrained RL, and dynamic system entification to perform reliably in real-world games. According to Yuntao Ma, a researcher at ETH Zurich, the controller’s end-to-end training optimizes the robot’s limbs for coordinated whole-body motion.

“We Brought the Future Back to Earth”: SpaceX Dragon Returns With 6,700 Lbs of Cutting-Edge Robotic Tech From Orbit

Robot Rallies Humans

The ANYmal-D robot has been tested against human players, demonstrating its ability to navigate the court and return shots at various speeds and angles. The robot achieved rallies of up to 10 consecutive hits, showcasing its capability to integrate whole-body movement with visual perception. By adjusting its gait based on timing and distance, the robot effectively tracks and intercepts shuttlecocks traveling at speeds of up to 40 feet per second. Impressively, the robot can rise onto its hind legs to keep the shuttlecock in view, prioritizing balance and safety to avoid falling.

“Your TV Has a Hidden Treasure”: This Free USB Port Unlocks 5 Little-Known Features That Instantly Boost Everyday Life

Despite these impressive feats, the robot faces challenges when competing against fast or aggressive shots, such as smashes. The team attributes the lower success rate in these scenarios to hardware limitations in camera perception and actuator speed, rather than deficiencies in the control algorithm. The framework’s adaptability to other sports or tasks is noteworthy, with researchers already extending it to robotic throwing tasks. This capability highlights the potential for integrating active perception into RL training loops, broadly applicable to tasks where perception and control must be tightly coordinated.

Future Directions and Enhancements

Looking ahead, improvements in perception responsiveness are essential for enabling longer rallies and full-court competitive gameplay. Currently, the system experiences an average delay of 0.375 seconds between the opponent’s shot and the robot’s first swing command. Reducing this latency is crucial for enhancing the robot’s ability to intercept faster and more distant shots, thereby enabling more extended rallies and competitive performance. Potential solutions include faster cameras or additional sensing modalities, which would significantly improve the robot’s capabilities.

The framework’s flexibility for generalization suggests promising applications beyond badminton, making it a template for deploying legged manipulators in other dynamic tasks. As the technology continues to advance, the possibilities for integrating robotics into various sectors are expanding, offering exciting prospects for the future of automated systems.

This advancement in robotics not only exemplifies the fusion of intelligence and machinery but also raises intriguing questions about the future of sports and automation. Can these developments lead to new forms of entertainment or even redefine how we perceive athleticism in the modern world?