Creative Adaptation by Evolution – CreAdapt

Evolutionary algorithms are good candidates to design an adaptive process because they are search processes that are both open and versatile. Nonetheless, current algorithms require several thousands of tests to obtain interesting solutions. To decrease the number of tests, the robot embeds a «self-model« -- a simulation of itself -- that allows it to execute most of the search in simulation; to constrain the search to behaviors that work on the real robot as well as in simulation, we rely on the «transferability approach«, which we introduced in previous work.

We will investigate this adaptation with wheel-legged hybrid robots because such robots can move in many different ways: they can use their wheels if the ground is flat, walk if their wheels are not working in the current environment, use a subset of their legs if one of them is broken, crawl if only a few degrees of freedom are working, ... Robots designed in the Creadapt project will therefore have a lot of possibilities when they will have to discover a new way to cope with a completely unforeseen situation.

Expected results are:

- new adaptation algorithms, both creative and fast;

- new methods to use artificial evolution for robots' locomotion;

- fundamental advances in the understanding of the evolutionary process that we observe in nature, and, in particular, about what makes organisms so evolvable (this will allow us to apply this new knowledge to design better algorithms).

This project aims at making robots more adaptive and more robust. The results could be applied in a «robotics rescue« scenario: when a robot has to move in an environment after a natural disaster, it is impossible to predict all the situations that the robot will encounter. Moreover, these are dangerous environment in which the robot can easily be damaged, for instance after a rock fall. In such a situation, it is critical to have robots that can continue their mission without the need of any human intervention.

Several publication are submitted.

The diversity and overall efficiency of lifeforms demonstrates the power of evolution as an optimization and innovation process. Evolutionary Algorithms (EAs) takes inspiration from this success and they are now mature black-box optimization algorithms and creative, automatic open-ended design methods.

Natural evolution is however more than a creative optimization process, it is also an efficient mechanism to adapt lifeforms to new environments and constraints. Importantly, such an adaptation to new situations is one of the main open challenges in robotics because autonomous robots have to deal with many situations unforeseen by their creators. Nevertheless, researchers in Evolutionary Robotics (ER) almost exclusively focused their efforts on the optimization abilities of the evolutionary process. Only a few, disparate works investigated the adaptive power of evolution for robotics.

The overall purpose of this project is to fulfill this gap by employing both the creative and the adaptive abilities of evolutionary algorithms to power software that can autonomously and creatively adapt the behavior of robots to unforeseen situations. In the typical scenario, a mobile robot faces a situation that requires adaptation (e.g. a leg is broken or the ground surface changed). The robot is allowed to launch a few experiments to investigate the situation; after a few minutes it should be able to pursue its mission until a new adaption is required. Algorithms that will be developed in this project will be tested on versatile wheel-legged robots based on commercially available ``bioloid'' kits.

Instead of trying to simulate the evolution of an ecosystem of robots (i.e. directly copy natural evolution), we will take inspiration from direct policy search algorithms that have been designed for reinforcement learning scenarios in robotics. To expand the range of possible behaviors and thus enable a creative adaptation, we will substitute the local optimization algorithm that power these algorithms with evolutionary ones because they can optimize both the structure and the parameters of controllers.

The main challenge to harness the creative and adaptive power of EAs in this fashion is to minimize the number of trials on the robot while allowing a very large (possibly infinite) number of behaviors. We will tackle this challenge with two complementary lines of research:
- taking inspiration from Bongard's work on self-model based resilient machines and from our previous work on the ``reality gap'', we will design a new adaptation scheme in which most of the optimization will rely on a simulated self-model while the search will be focused on controllers that work on the real robot in the current situation;
- we will improve evolvability (i.e. the ability to quickly adapt to new situations) as much as possible by using a bio-inspired encoding (HyperNEAT + non-linear oscillators) and by encouraging modularity (by adding new selective pressures).

Overall, our project aims at improving the state of the art in autonomous robotics by making fundamental advances in adaptation algorithms. The support of the ANR will allow the building of three robotic demonstrators and the hiring of a 2 years post-doctoral fellow; it will mainly result in scientific publications and open source software.