Modeling Social Common Sense for Seamless Human Machine Teaming: Inverting the Intuitive Game Engine with Probabilistic Programming
Abstract
Humans use commonsense knowledge of agents and the physical world to solve problems interactively. We model peoples knowledge of agents and the world using rich game-engine-based models of 3D physical scenes with realistic physical dynamics and agents autonomously acting and interacting with others, based on individual mental states and shared tasks. Cast as a probabilistic program, this intuitive game engine can be inverted to support inferences about others beliefs, desires, goals, and tasks, which are vital for successful social interaction. We show the promise of this approach in application to modeling human inferences of the targets of the observed reaching actions of others.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jul 06, 2017
- Accession Number
- AD1059216
Entities
People
- Chris Baker
- Emanuel Todorov
- Joshua B. Tenenbaum
- Tao Gao
- Vikash Kumar
Organizations
- Massachusetts Institute of Technology