Instead of the current method of character motion capture which makes use of animators or actors and motion capture cameras, the DeepLoco algorithm utilizes a type of cutting-edge machine learning algorithm called deep reinforcement learning. The DeepLoco software which Profesor van de Panne will present at SIGGRAPH 2017, learns by experience through incentive based trial and error to continuously recognize better movement strategies to take in specific situations. The end result is the characters are enabled to move in realistic ways in harmony with their environment, with examples achieved (See short video below) such as; characters that learn to successfully walk along a narrow path while avoiding running into people or other moving obstacles, or even creation of a sports character to dribble a soccer ball towards a goal. DeepLoco might be the future of video, movie and cartoon animation. Profesor van de Panne also sees DeepLoco as a better way to self-teach robots to navigate through their environment instead of needing to hand-code the appropriate rules.
Breakthrough Deep Reinforcement Learning Software Trains Computer Characters to Realistically Move
DeepLoco, which offers a simpler fashion to animate human motion in games and film.
Anyone that has worked with animation software, knows what a challenge it is to make characters move naturally. Michiel van de Panne, a University of British Columbia (UBC) computer science professor has created a new software called