Learning to Swing

2018 - Computer Animation Class Project

Project Description

This Computer Animation class project utilized off-the-shelf Soft-Actor-Critic (SAC) Reinforcement Learning method to teach an animated character to build up momentum and swing on a pull-up bar.

The project demonstrates how modern RL algorithms can be applied to character animation tasks, creating realistic and dynamic movement patterns that emerge from the learning process rather than being hand-crafted.

Technical Approach

  • • Soft Actor-Critic (SAC) reinforcement learning algorithm
  • • Physics-based character simulation
  • • Reward shaping for momentum building and swinging behavior
  • • Real-time character animation from learned policies
  • • Integration with physics simulation environment

Results

The trained agent successfully learned to generate realistic swinging motions, demonstrating the ability to: build up initial momentum, maintain rhythmic swinging patterns, and adapt to different starting conditions. The resulting animations showed natural-looking character movement that would be difficult to achieve through traditional keyframe animation techniques.