Quadrupedal Robot Dog, Now Playing Badminton? A Leading Case of Physical Intelligence (PI) Machine Learning

Project Overview

This implementation case demonstrates advanced Physical Intelligence (PI) based sports robot technology realized through the quadruped robot ANYmal-D developed at ETH Zurich, Switzerland. ANYmal-D is a cognitive robot capable of engaging in actual badminton rallies with humans through reinforcement learning. It recognizes the trajectory of the shuttlecock using only on-board sensors without external sensors and performs agile plays by moving strategically using its entire body. This project is drawing attention as a practical demonstration of situational responsive automation technology, where robots judge and react independently in dynamic environments, going beyond simple automation.

Project Background & Objectives

Background

Conventional autonomous robots typically had limitations in performing repetitive tasks or following fixed trajectories in static environments. However, in unpredictable situations such as sports, advanced automation technology capable of real-time perception, judgment, and motion control is required. The research team at ETH Zurich pursued the ANYmal-D project with the goal of developing a robot system equipped with human-level athletic intelligence.

Objectives

Implementing autonomous responsive motion control in unstructured environments
Developing an integrated perception-motion system using on-board sensors
Technical demonstration of sports-type robots capable of collaborating with humans
Testing reinforcement learning algorithms based on Physical Intelligence (PI)

Components

Robot	Robot ANYmal-D: A cognitive robot combining quadruped-based mobility with upper-body manipulation functions Robot Arm Actuator: A dedicated motion unit capable of holding a racket and striking

Workflow

STEP 1.	Detect shuttlecock and identify based on color using on-board sensors.
STEP 2.	Predict the trajectory of the shuttlecock and calculate the target landing point.
STEP 3.	Select movement strategy (Stop/Run/Jump, etc.).
STEP 4.	Counterattack the shuttlecock by swinging the racket with appropriate timing and posture.
STEP 5.	Automatically return to the center of the court after striking and prepare for the next move. ※ Notice: Unauthorized copying or recreation of any content within Marosol may violate the Unfair Competition Prevention Act and the Copyright Act.

※ The content provided by Marosol is protected under U.S. and Canadian copyright and intellectual property laws. Unauthorized reproduction, distribution, or use of this content is strictly prohibited.

Key Features

High-Dimensional Perception-Based Automation A robot system capable of real-time situational judgment and strategy formulation, rather than simple repetitive motions. Implementation of a reinforcement learning model that enables trajectory prediction even within limited visibility and sensor noise.
Full-Body Physical Intelligence Implementation The robot performs agile movements by organically combining its torso, legs, and arms. Autonomous selection of the optimal movement strategy based on the shuttlecock's position, speed, and airtime.
Autonomous Return and Preparation for Next Play Implementation of actions that include moving to the center of the court after striking to prepare for the next attack. An advanced control system that mimics the human-like motion flow required in sports competitions.

Implementation Results

ROI Period
Key Metrics
Client Feedback	The ETH Zurich research team stated, "ANYmal-D has surpassed the limitations of conventional robots by implementing true Physical Intelligence that perceives and reacts to physical situations." They evaluated it as "a turning point that proves the potential for robots that move and react like humans in various fields such as sports, disaster rescue, and military training."