Real-time Diverse Motion In-betweening with Space-time Control

TL;DR

This work presents a data-driven framework for generating diverse in-betweening motions for kinematic characters that injects dynamic conditions and explicit motion controls into the procedure of motion transitions, enabling rich, versatile, and high-quality animation generation.

Abstract

In this work, we present a data-driven framework for generating diverse in-betweening motions for kinematic characters. Our approach injects dynamic conditions and explicit motion controls into the procedure of motion transitions. Notably, this integration enables a finer-grained spatial-temporal control by allowing users to impart additional conditions, such as duration, path, style, etc., into the in-betweening process. We demonstrate that our in-betweening approach can synthesize both locomotion and unstructured motions, enabling rich, versatile, and high-quality animation generation.

Figures (5)

• Figure 1: The DC-MoE framework consists of two parts. The gating network takes the condition that contains the phases predicted by the network in an autoregressive manner, along with style and the time-to-arrive at the target pose as inputs, and outputs gating weights. The expert network takes the current pose state $P_{i}$, target pose state $P_{target}$, along with the root trajectory centered at $P_{i}$ as inputs, and gets blended with the gating weights to output the pose features in the next timestamp.
• Figure 2: Trajectory of the character's root produced by arbitrary route planning with different root orientations of the start and target pose under same start-to-end distance. The exploration in the RTN latent space does not produce varied trajectories and tends to collapse into an identical outcome. By utilizing random sampling from the trajectory gallery, our method allows the character to explore a wider range of in-between trajectories and diverse time durations.
• Figure 3: Visualizing the distance distribution across different styles. Given the $45$ frames segments, we show the cumulative trajectory within this segment. Lighter colors denote fewer instances of a particular motion style, while darker colors indicate more numerous instances, which show the distribution of different motion styles.
• Figure 4: Example of with and without using style condition, while all other input features are kept same. The results indicate that with a style condition, the model can generate more dynamic motion.
• Figure 5: Example of matching the candidate trajectory from the gallery with the same start-to-end distance as the input query.