Combining Recurrent Neural Networks and Adversarial Training for Human Motion Synthesis and Control.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Human motion synthesis and control are crucial for realistic character animation and virtual environments.
• Existing methods often struggle with long-term temporal dependencies and generating natural, diverse motions.

Purpose of the Study:

• To introduce a novel generative deep learning network for human motion synthesis and control.
• To develop a compact, contact-aware model capable of generating infinite, natural-looking motions.

Main Methods:

• Utilized recurrent neural networks (RNNs) with long short-term memory (LSTM) cells for modeling nonlinear dynamics.
• Employed adversarial training, inspired by generative adversarial networks (GANs), for refining motion sequences.
• Trained a discriminative network to ensure refined motions are indistinguishable from real motion capture data.

Main Results:

• The proposed deep learning model generates highly realistic human motions comparable to high-quality motion capture data.
• The model is compact, contact-aware, and capable of producing infinitely long, natural motion sequences.
• Experiments demonstrated superior performance against baseline models in various applications.

Conclusions:

• The novel generative deep learning network effectively synthesizes and controls human motion.
• The combination of RNNs and adversarial training offers a powerful approach for realistic motion modeling.
• The model shows significant potential for applications including random motion synthesis, online/offline control, and motion filtering.