Toward the Generation of Natural Animal-like Behavior: Motion Generation Methods and Evaluation

—Approach Using gMLP with Large-scale Video Data —

Abstract

This study aims to create a robot capable of diverse behaviors. To achieve this, we propose a behavior generationframework for robots and evaluate whether the generated behaviors result in diverse movements. The proposedframework learns cat behaviors, which are familiar to humans, through a large language model (LLM). The datasetwas created by extracting 2D keypoints from cat videos, converting these 2D keypoints into 3D keypoints, and calculatingjoint angles using inverse kinematics based on the joint positions. The data, consisting of joint positions andangles, was then converted into a language-like format, called ”motion language.” This converted data was used totrain the gMLP. The gMLP’s loss function reached its minimum validation loss after two epochs, so the model trainedup to this point was used to generate motion language. Evaluation experiments assessed whether the generated behaviorsequences could sustain diverse behaviors over time. The evaluation method used Multivariate Multiscale Entropy.The results confirmed that the gMLP-generated data exhibited a level of complexity comparable to that of the trainingdata. Additionally, the generated data demonstrated greater complexity than random data. gMLP’s output not onlyachieved short-term complexity but also generated long-term sequences with a level of complexity unattainable byrandom behavior generation.