Abstract
Conventionally, robot morphologies are developed through simulations and calculations, and different control methods are applied afterwards. Assuming that simulations and predictions are simplified representations of our reality, how sure can roboticists be that the chosen morphology is the most adequate for the possible control choices in the real-world? Here we study the influence of the design parameters in the creation of a robot with a Bayesian morphology-control (MC) co-optimization process. A robot autonomously creates child robots from a set of possible design parameters and uses Bayesian Optimization (BO) to infer the best locomotion behavior from real world experiments. Then, we systematically change from an MC co-optimization to a control-only (C) optimization, which better represents the traditional way that robots are developed, to explore the trade-off between these two methods. We show that although C processes can greatly improve the behavior of poor morphologies, such agents are still outperformed by MC co-optimization results with as few as 25 iterations. Our findings, on one hand, suggest that BO should be used in the design process of robots for both morphological and control parameters to reach optimal performance, and on the other hand, point to the downfall of current design methods in face of new search techniques.
Highlights
As robots are, in their final form, physical representations of the model they were designed to be, discrepancies between design tools and reality augur poorly on their final behavior
We present a design process where a robot creates another robot by choosing MC design parameters while inferring its real-world behavior
Unlike current robotic applications with machine learning, where the control parameters are the only ones to be optimized in the real-world [3,4,5], this work introduces the idea of machines intelligently and purposefully designing their body and mind to reach a better real world behavior
Summary
In their final form, physical representations of the model they were designed to be, discrepancies between design tools and reality augur poorly on their final behavior. This difference, known as the Reality Gap [1,2], can be reduced with programming methods to improve control parameters within the robot, but morphological parameters are difficult to modify and rarely changed, which incurs in a sub-optimal behavior. We present a design process where a robot creates another robot by choosing MC design parameters while inferring its real-world behavior This robot uses BO to predict the influence of the chosen morphological and control parameters on the behavior of its child, and iteratively changes both parameters to find the best MC pair. Unlike current robotic applications with machine learning, where the control parameters are the only ones to be optimized in the real-world [3,4,5], this work introduces the idea of machines intelligently and purposefully designing their body and mind to reach a better real world behavior
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
