Morphology dictates learnability in neural controllers

Abstract

Catastrophic forgetting continues to severely restrict the learnability of controllers suitable for multiple task environments. Efforts to combat catastrophic forgetting reported in the literature to date have focused on how control systems can be updated more rapidly, hastening their adjustment from good initial settings to new environments, or more circumspectly, suppressing their ability to overfit to any one environment. When using robots, the environment includes the robot's own body, its shape and material properties, and how its actuators and sensors are distributed along its mechanical structure. Here we demonstrate for the first time how one such design decision (sensor placement) can alter the landscape of the loss function itself, either expanding or shrinking the weight manifolds containing suitable controllers for each individual task, thus increasing or decreasing their probability of overlap across tasks, and thus reducing or inducing the potential for catastrophic forgetting.

Living Things Are Not (20th Century) Machines: Updating Mechanism Metaphors in Light of the Modern Science of Machine Behavior

“…However, these methods can be seen as deepening rather than narrowing the brain/body distinction: In these approaches, the robot's form is usually a fixed shell, previously designed by human engineers, controlled by the machine learning algorithm. In contrast, there is a small but growing literature on embodying intelligence directly into the body of the robot (Nakajima et al, 2015), and in machine learning methods that evolve robot bodies to enhance this and other forms of intelligence (Powers et al, 2020). A small but growing literature on robots capable of self-modeling also blurs the distinction between embodied robots and non-embodied AI methods.…”

Beyond Body Shape and Brain: Evolving the Sensory Apparatus of Voxel-Based Soft Robots

“…To the best of our knowledge, the only work that addressed a similar research question are [16][17][18], where the sensory apparatus is evolved in different kinds of hard robots. Another work on this topic is [19], which has shown that sensor placement can alter the landscape of the controller loss function, thus guiding evolution towards better controllers. However, the context of that work was unicycle non-holonomic mobile robots.…”

Optimizing the Sensory Apparatus of Voxel-Based Soft Robots Through Evolution and Babbling

“…To the best of our knowledge, the only works that addressed similar research questions are [19][20][21], although those works focused on evolving the sensory apparatus of hard robots. Another related work is [22], where it has been shown that the landscape of the controller loss function depends on the sensors placement, which in turn conducts evolution towards better controllers. However, the context of that work is unicycle non-holonomic mobile robots.…”