Bioinspired Dynamic Control of Amphibious Articulated Creatures with Spiking Neural Networks

Ioannis Polykretis, Mridul Aanjaneya, Konstantinos P. Michmizos

Keywords: spiking neural networks, articulated characters, locomotion control, physics simulation

TL;DR: We present a biologically plausible, compact spiking neural network for controlling the crawling and swimming behaviors of amphibious creatures with articulated skeletons.

Abstract: We present a biologically plausible, compact spiking neural network for controlling the crawling and swimming behaviors of amphibious creatures with articulated skeletons. Prior methods for learning efficient control policies for such creatures are resource-greedy, both in terms of computational time and energy requirements due to the high number of degrees of freedom introduced by the many joints present in the creature skeleton. Our approach takes a radical departure from prior work and exploits the physiology of amphibious creatures. Specifically, we emulate experimentally identified biological controllers for amphibious creatures with a network of spiking neurons, which alleviates the need for training altogether and can potentially provide the additional benefit of utilizing minimal resources in terms of energy. Our approach is robust and allows the amphibious creature to avoid both static and dynamic obstacles when exhibiting different movement patterns, and also adaptively control its swimming speed. Moreover, we show that the creature can seamlessly transition between crawling and swimming behaviors as it moves from land to water or vice-versa, similar to its real-world counterpart. Our approach presents an efficient and scalable alternative for producing vivid and lively motion, as we demonstrate through a complex scene where multiple amphibious creatures interact with each other, successfully avoiding collisions while moving across a pool of water. Our approach is generalizable to other creatures also, as we show through the design of a controller for a quadruped.

Track: Graphics

Summary Of Changes:

Author response for changes
We thank the reviewers for their feedback and have addressed all points raised by them. Below is the list of the main changes in our manuscript in response to the reviewers’ comments and the related pointers to the revised manuscript. The new text in the revised manuscript is marked in red.

A major point that was brought up by all reviewers was that the swimming and crawling tasks of our agent were less challenging than the previously demonstrated biped locomotion, which would require balancing based on extensive sensory feedback. We now extend our discussion about the limitations of our method (Section 9, paragraph 2) to include this point. We also add this direction as possible Future Work (Section 10, paragraph 2) with the inclusion of more and increasingly sophisticated sensors.

Given the concern raised by the reviewers regarding the comparison of our method against Reinforcement Learning, we now expand our comparison discussion. More specifically, we modify Section 8.9 to better reflect that the comparison is not completely fair: While our method does not require thousands of training episodes to learn the behaviors from scratch, it does require domain expert knowledge of biological network architectures. We also tone down our “no-training” claim in our Introduction (Section 1, last paragraph) to better represent the benefits and drawbacks of our method.

To discuss the possibility of using simpler control methods (i.e., sinusoidal functions) that was brought up, we extend our discussion of Related Works (Section 2, paragraph 3). There, we clarify that while the repetitive behavior indeed consists of a sinusoidal trajectory of each joint, a simple sinusoidal function would not be sufficient to handle the control task. The reason for this is the lack of adaptability of such a controller: Changes in the frequency or the phase of the sinusoids cannot be addressed by static function generators and require at least some dynamical system behavior. The introduction of such oscillators and their drawbacks are already discussed in our Related Work Section (Section 2, paragraph 3).

To address the concern about our argument regarding the energy efficiency of our method, we add a brief discussion of the estimated energy consumption (Section 4, last paragraph). We agree with the reviewers that the power required for the motors is generally dominating the overall system consumption. However, our method targets the computational aspects of control; hence, a fair comparison can be made against other methods with the same goal, i.e., planning the motor activations, integration of sensory feedback and behavior adaptation, etc. In that domain, the use of spiking neurons with minimal training requirements does provide considerable improvement in energy consumption.