In a remarkable leap for robotic dexterity, researchers at ETH Zurich have trained their quadrupedal robot ANYmal to accurately pick up and hurl objects like a human athlete. This advancement transforms the machine from a mobile inspection platform into a versatile manipulator capable of dynamic interactions in unpredictable environments. Equipped with a custom robotic arm featuring six degrees of freedom, ANYmal can now grasp diverse objects from tennis balls to fruit and propel them toward targets up to five meters away, even while navigating challenging terrain like gravel, grass, or concrete.
The breakthrough stems from a sophisticated reinforcement learning framework combined with sim-to-real transfer technology. Researchers, led by Fabian Jenelten of ETH Zurich’s Robotic Systems Lab, first immersed ANYmal in a hyper-realistic virtual environment. Within this digital sandbox, the robot practiced millions of throws without risk of damage, while physics engines meticulously replicated real-world conditions like gravity, wind resistance, and surface interactions. “Simulation’s a game-changer,” Jenelten explains. “You can try endless scenarios, but the real win is making those lessons hold up when the robot steps outside”.
Engineering Adaptive Intelligence
Central to ANYmal’s newfound skill is its ability to autonomously adjust its stance and force based on shifting conditions. When lifting an object, the robot dynamically redistributes its weight to maintain balance. During throws, it fine-tunes motion trajectories to compensate for variables like wind gusts or uneven footing capabilities that stymie traditional robots reliant on pre-programmed sequences. “Robustness was our focus,” emphasizes researcher Joonho Lee. “This robot has to deal with different weights, shapes, even a gust of wind, and still get it right”. Unlike factory robots operating in controlled settings, ANYmal thrives in chaos, a trait enabled by compliant actuators delivering precise torque control for dynamic stability.
Beyond Novelty: Toward Real-World Utility
This development transcends mere spectacle. It represents a critical stride toward general-purpose robotic autonomy in high-stakes scenarios like disaster response or industrial inspection. ANYmal’s core design targets operation in environments hostile to humans—offshore oil rigs, earthquake-damaged structures, or nuclear facilities—where it already navigates stairs, obstacles, and slippery surfaces using integrated LIDAR and cameras for 3D mapping. Adding sophisticated manipulation unlocks new capabilities: clearing debris, handling tools, or deploying sensors. Marco Hutter, Professor of Robotics, underscores this trajectory: “We wanted to see how far we could push a legged robot”.
The integration of learning-based control with robust mobility hints at a future where robots adapt on the fly to novel challenges. ETH Zurich’s approach diverges from conventional robotics by prioritizing resilience through AI-driven trial-and-error, akin to how a child learns. This philosophy is further evidenced in ANYmal’s locomotion systems, where reinforcement learning in simulation produced controllers capable of “blind” traversal across mud, snow, and rubble, surpassing optimization-based methods.
The Sim-to-Real Advantage
The project underscores the accelerating role of simulated training environments in robotics. By bridging the “reality gap” through techniques like domain randomization, which vary virtual textures, lighting, and physics during training, researchers achieve unprecedented transferability to physical hardware. A recent survey in IEEE Robotics and Automation Letters confirms this methodology’s growing dominance for complex mobile manipulation tasks. For ANYmal, thousands of synthetic throws translated directly into real-world accuracy, proving that virtual rehearsal, when engineered meticulously, equips robots to handle life’s unpredictability. As Jenelten observes: “The real world’s messy… Most robots trip over that, but ANYmal? It thrives”
Subscribe to my whatsapp channel
Comments are closed.