Reinforcement Learning for Animation Character Bone Training

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current techniques for creating animated characters in video games are time-consuming and computationally expensive, as they often rely on reusing a single skeleton or employing motion retargeting, which can fail when characters have vastly different structures.

Innovation Solution

The use of reinforcement learning to train animation characters by mapping disparate target animation data to specific parts of a bone hierarchy, allowing for the blending of diverse motion sequences within a simulation environment, thereby creating realistic movements.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If motion retargeting is used to transform motion to a new character, then the motion can be customized to the new character, but the process is still time consuming and computationally expensive

Engineering Contradiction:
Improvemotion customizationVSAvoidtraining time
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent divides the character skeleton into multiple bone hierarchies (e.g., left arm, right arm, left leg, right leg, torso) and trains each bone hierarchy separately using reinforcement learning. This segmentation allows parallel training of different body parts, significantly reducing overall training time while maintaining customized motion for each character.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent pre-trains bone hierarchies using reinforcement learning to learn optimal motion policies before actual animation generation. This preliminary training phase allows the system to quickly generate customized motions for new characters without performing time-consuming iterative joint manipulation at runtime.

Inventive Principle:
Principle #10Preliminary action

2Adaptability or versatility

If motion retargeting uses joint manipulation to generate transformed motion, then the motion can be customized, but the process is computationally expensive

Engineering Contradiction:
Improvemotion customizationVSAvoidcomputational resources
Core Design Contradiction:
Adaptability or versatilityVSUse of energy by moving object

Solution Approach 1:

The patent replaces the traditional mechanical joint manipulation approach with reinforcement learning-based policy networks. Instead of iteratively adjusting joints through computational optimization, the pre-trained neural networks directly predict bone hierarchy transformations, dramatically reducing computational resource requirements while maintaining motion customization capability.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The computationally intensive training process is performed in advance during an offline phase, creating pre-trained policy networks. During actual animation generation, the system only needs to execute the pre-trained networks, which requires minimal computational resources compared to real-time joint manipulation.

Inventive Principle:
Principle #10Preliminary action

3Productivity

If a single skeleton is reused for multiple characters, then character creation time is reduced, but all characters will look alike and have the same movements

Engineering Contradiction:
Improvecharacter creation speedVSAvoidcharacter diversity
Core Design Contradiction:
ProductivityVSAdaptability or versatility

Solution Approach 1:

The patent segments the character skeleton into multiple independent bone hierarchies that can be independently trained and transformed. This allows the same base skeleton to generate diverse character movements by applying different transformations to each bone hierarchy, achieving character diversity while maintaining efficient character creation.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies different motion transformations to different parts of the skeleton (e.g., different arm movements, leg movements, torso movements) based on character-specific parameters. This local customization allows characters to have unique movement patterns while sharing the same overall skeleton structure, achieving both efficiency and diversity.

Inventive Principle:
Principle #3Local quality

4Adaptability or versatility

If motion retargeting is used to transform motion, then motion can be adapted to new characters, but it may fail when the new character is vastly different than the base character

Engineering Contradiction:
Improvemotion adaptationVSAvoidmotion transformation success
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent uses reinforcement learning to learn robust motion policies that can adapt to various character parameters and proportions. The trained policy networks generalize well to characters with vastly different structures from the training data, ensuring reliable motion transformation even when the new character differs significantly from the base character.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent creates universal policy networks through reinforcement learning that can handle multiple character types and motion scenarios. These trained networks are not limited to specific character proportions or motion styles, making the motion adaptation reliable across diverse character designs and movement patterns.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP3939010B1Reinforcement learning to train a character using disparate target animation data
Publication Date: 2024.04.03 SONY INTERACTIVE ENTERTAINMENT LLC
  • EP3939010B1 patent drawingFigure 1A
  • EP3939010B1 patent drawingFigure 1B
  • EP3939010B1 patent drawingFigure 2

AI summary

A method for training an animation character, including mapping first animation data defining a first motion sequence to a first subset of bones of a trained character, and mapping second animation data defining a second motion sequence to a second subset of bones. A bone hierarchy includes the first subset of bones and second subset of bones. Reinforcement learning is applied iteratively for training the first subset of bones using the first animation data and for training the second subset of bones using the second animation data. Training of each subset of bones is performed concurrently at each iteration. Training includes adjusting orientations of bones. The first subset of bones is composited with the second subset of bones at each iteration by applying physics parameters of a simulation environment to the adjusted orientations of bones in the first and second subset of bones.