ETH Zurich, through its Robotic Systems Lab (RSL), continues to redefine robotic mobility with its latest advances in AI-controlled legged robotics. The ANYmal-D quadrupedal robot, developed by ANYbotics and enhanced by ETH’s AI-driven control architecture, showcases state-of-the-art mobility intelligence, terrain adaptability, and real-world autonomy.
By fusing reinforcement learning, model predictive control, and vision-based terrain awareness, ETH Zurich’s AI system pushes ANYmal-D closer to the ultimate goal of fully autonomous, intelligent robots for use in industrial, inspection, and search-and-rescue scenarios.
ANYmal is a quadrupedal robot designed for autonomous operation in challenging environments. Driven by special compliant and precisely torque controllable actuators, the system is capable of dynamic running and high-mobile climbing. Thanks to incorporated laser sensors and cameras, the robot can perceive its environment to continuously create maps and accurately localize. Based on this information, it can autonomously plan its navigation path and carefully select footholds while walking. Driven by our first real-world application, namely industrial inspection of oil and gas sites, ANYmal carries batteries for more than 2h autonomy and different sensory equipment such as optical and thermal cameras, microphones, gas-detection sensors and active lighting. With this payload, the machine weighs less than 30kg and can hence be easily transported and deployed by a single operator.
The ANYmal robot is available through external pageANYbotics. For technical details, check out the ANYmal Website. For further information, please contact info@anybotics.com.
Overview of ANYmal-D Robot Platform
- Form Factor: Quadruped robot (similar in morphology to a medium-sized dog).
- Actuation: Custom high-torque series elastic actuators (SEA) in each joint.
- Payload: Equipped with LIDAR, stereo/depth cameras, IMU, and onboard GPU/CPU.
- Power: Battery-powered with swappable modules, enabling 2–4 hours of operation.
- Applications: Industrial inspection, underground tunnels, offshore rigs, disaster zones.
AI Controller Architecture
ETH Zurich’s AI controller architecture for ANYmal-D integrates machine learning with real-time planning and reactive control, forming a hybrid intelligence stack:
1. Reinforcement Learning Policy (Offline Trained)
- Trained in simulation using Proximal Policy Optimization (PPO).
- Enables locomotion strategies like walking, trotting, climbing stairs, jumping.
- Policies generalized over thousands of terrain configurations (rocks, slopes, gaps).
2. Terrain-Aware Perception
- Fuses input from:
- Stereo cameras
- Depth sensors
- LIDAR
- Builds a heightmap and traversability cost map of the terrain in real time.
3. Model Predictive Control (MPC) Layer
- Optimizes over 100–300 ms horizon to ensure dynamic stability and foot placement.
- Incorporates torque limits, friction coefficients, and body pose optimization.
4. Behavior Selector (Hierarchical AI)
- Switches between motion primitives (walk, step over, crawl) based on terrain confidence.
- Utilizes a supervisory neural net trained with human-labeled terrain data.
Research Collaborators and Contributions
- ETH Zurich RSL: AI locomotion, terrain learning, MPC.
- ANYbotics AG: Mechatronics, embedded systems, industrial testing.
- NVIDIA & Isaac Sim: High-fidelity training environments and parallel simulation pipelines.
- DARPA & ESA: Funding and deployment environments for autonomous legged robots.
Swiss Researchers Develop AI-Powered Robot That Plays Badminton Against Humans
Using reinforcement learning, the robot accurately tracks the shuttlecock’s flight, predicts its trajectory, and navigates the court to intercept and return shots.
Researchers at a Switzerland university have developed an AI-powered legged robot that plays badminton against humans with impressive agility. Researchers at ETH Zurich tested their AI controller on ANYmal-D, a four-legged robot equipped with a stereo camera and a dynamic arm holding a badminton racket.