Sites Inria

Version française

Research

27/05/2015

Robots that adapt to damage in a few minutes

Robots could help our society in many types of situation, for example searching for survivors after a natural disaster or alerting the fire services in the event of a forest fire. But they will always be confined to research laboratories until they are able to continue operating when damaged. Researchers from the Institute for Intelligent Systems and Robotics (CNRS/UPMC) and the Lorraine Laboratory of Research in Computer Science and its Applications (CNRS/Inria/University of Lorraine) are showing how robots can automatically adapt to damage in less than two minutes. Their results were published in the 28 May 2015 issue of Nature.

Unlike present-day robots, living beings have an impressive ability to adapt to injury. For example, most amputee dogs are able to play, jump, and run despite missing a paw, and a child with a sprained ankle only takes a few minutes to find a way of limping on it. The researchers took inspiration from these examples. "When animals are injured, they are not completely unequipped to cope with it," explains Jean-Baptiste Mouret. "On the contrary: they have good instincts about different ways of reacting. This intuition helps them make an intelligent choice of behaviours to try, and after testing a few things, they find something that works despite the injury. Our robots do the same thing."

Before being sent out on a mission, the robot uses a simulation of its body to create a detailed "map" of the thousands of different ways it could perform its task. This map represents the robot's intuition concerning the behaviours of interest and their potential. If the robot is damaged, it uses its intuition to guide a learning algorithm which conducts experiments to find a compensatory behaviour quickly. The new algorithm is called "Intelligent Trial and Error".

 "If damaged, our robot behaves like a scientist," explains Antoine Cully. "It has prior knowledge about the various actions that could work, and it starts by testing them. However, the prior knowledge comes from the simulation of the intact robot. It must therefore find the approaches that work not only in reality, but also with the damage.Every action it tests is like an experiment to confirm or rule out its hypotheses. If an action does not work, the algorithm intelligently eliminates entire categories of action and tries completely different things. For example, if walking with the weight almost entirely on the rear legs does not work well, the robot will try walking with its weight on the front legs. It is surprising how quickly the robot finds a new way of walking. Despite a leg cut in half, it only takes the robot two minutes to find an efficient way of limping!"

Jeff Clune explains: "Technically, the algorithm is divided into two stages: (1) creating the map of performance behaviour space, and (2) adapting to the new situation."

The map from the first stage is created with a new type of evolutionary algorithm called MAP-Elites. This type of algorithm is inspired by Darwinian evolution and "survival of the fittest" to find solutions that perform well. The adaptation in the second stage is based on a Bayesian optimisation algorithm which exploits the prior knowledge provided by the map to find a new behaviour quickly.

"We conducted experiments which show that the key point is in the creation and processing of the prior knowledge," continues Jeff Clune. This new technique could contribute to developing more robust, more efficient autonomous robots. Danesh Tarapore gives a few examples. "This could allow us to create robots capable of assisting rescue workers without requiring their constant attention," he says. "It could also facilitate the creation of robotic personal assistants that can continue being useful even when a part is broken."

For more information, a video illustrating this work: https://youtu.be/T-c17RKh3uE

A six-legged robot learns to walk again with one damaged leg and one missing leg. The experiment is repeated with a robotic arm learning to position an object correctly despite several jammed motors.

 

This work has received support from the National Research Agency (Creadapt, ANR-12-JS03-0009), the European Research Commission (ResiBots, grant agreement No 637972), and the Direction Générale de l’Armement (General Weapons Directorate) (thesis by A. Cully).

Keywords: CNRS ERC Robotique Algorithme UPMC Nature

Top