Evolving coordinated quadruped gaits with the HyperNEAT generative encoding

Abstract

Abstract-Legged robots show promise for complex mobility tasks, such as navigating rough terrain, but the design of their control software is both challenging and laborious. Traditional evolutionary algorithms can produce these controllers, but require manual decomposition or other problem simplification because conventionally-used direct encodings have trouble taking advantage of a problem's regularities and symmetries. Such active intervention is time consuming, limits the range of potential solutions, and requires the user to possess a deep understanding of the problem's structure. This paper demonstrates that HyperNEAT, a new and promising generative encoding for evolving neural networks, can evolve quadruped gaits without an engineer manually decomposing the problem. Analyses suggest that HyperNEAT is successful because it employs a generative encoding that can more easily reuse phenotypic modules. It is also one of the first neuroevolutionary algorithms that exploits a problem's geometric symmetries, which may aid its performance. We compare HyperNEAT to FT-NEAT, a direct encoding control, and find that HyperNEAT is able to evolve impressive quadruped gaits and vastly outperforms FT-NEAT. Comparative analyses reveal that HyperNEAT individuals are more holistically affected by genetic operators, resulting in better leg coordination. Overall, the results suggest that HyperNEAT is a powerful algorithm for evolving control systems for complex, yet regular, devices, such as robots.

Evolving neural fields for problems with large input and output spaces

“…In the particular setup reported, the outputs of the HyperNEAT network are used as desired joint angles for the robot, and a P controller is used to approach these angles. In addition, a sine wave is provided as input to the network to ease the evolution of regular gaits (Clune et al, 2009a). While performance on the multilegged walking task was best with a hand designed input and output geometry, HyperNEAT still outperformed a direct encoding when the geometry was randomized (Clune et al, 2009b).…”

Lexicase selection outperforms previous strategies for incremental evolution of virtual creature controllers

“…It has demonstrated an ability to address challenging problems in many task domains including: gaits (Clune et al, 2009), object manipulation (Bongard, 2008), biological study (Crespi et al, 2013;Doorly et al, 2009), and the optimization of morphology (Auerbach and Bongard, 2010;Bongard, 2010;Cheney et al, 2013). Often, these tasks have a single performance objective, or weighted sum, to assess the fitness of each individual.…”

HyperNEAT Versus RL PoWER for Online Gait Learning in Modular Robots

“…A gait control table consist of rows of actuator commands with one column for each actuator, each row also has a condition for the transition to the next row.A second major avenue of research is that of neural networks (NN). In particular for locomotion of robot organisms HyperNEAT is used extensively with several studies showing that HyperNEAT is capable of creating efficient gaits for robots [4,9,20]. HyperNEAT is discussed in more detail in Section 3.…”

“…These times were chosen because they are multiples of the sine wave period and were found to produce better results and avoid organism flipping because of too harsh transitions between gaits. The fitness of each controller F i is calculated as in the original experiment [4] and is defined as…”

“…In this paper we employ an evolutionary approach HyperNEAT [4] that has a good reputation for this type of tasks and compare it with PoWeR. Similarly to [5] we use the learning algorithms with parameter values as recommended by the authors.…”