Action-Based Models for Learned Task Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for training automated devices to recognize human behavior, such as picking up a cup or kicking a ball, often suffer from overlearning and performance degradation in novel situations, and fail to adequately account for physical phenomena like gravity and inertia.
Innovation Solution
A 'learning by discovery' method where an agent interacts with a simulated three-dimensional learning environment to execute and recognize actions, generating observations, hypothesizing action models, and vetting them to identify learned models capable of performing tasks, using a reinforcement learning framework and internal action models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If statistical approaches are used to train automated devices to recognize actions, then training can be performed with labeled datasets, but the device suffers from overlearning and performance degradation when confronted with novel circumstances
Solution Approach 1:
Instead of training the device to recognize actions through statistical pattern matching, the invention inverts the approach by training the device to perform actions and then inferring recognition capability from performance capability. The device learns action models by attempting to execute actions in a simulated environment, and recognition is achieved by evaluating whether observed actions match the learned models, thereby improving generalization to novel circumstances while maintaining ease of training through the same reinforcement learning framework
2Device complexity
If statistical methods are used for action recognition, then implementation is straightforward, but the methods are unable to account for physical phenomena such as gravity and inertia
Solution Approach 1:
The invention replaces statistical pattern recognition with a physics-based action model that incorporates mechanical principles. The device learns to predict future states by applying physics equations (such as gravity and inertia) to current states, creating action models that are inherently grounded in physical reality. This substitution maintains reasonable implementation complexity through standardized physics simulations while dramatically improving reliability in physical contexts by explicitly modeling gravitational and inertial effects
Data Source
AI summary
Systems provided herein include a learning environment and an agent. The learning environment includes an avatar and an object. A state signal corresponding to a state of the learning environment includes a location and orientation of the avatar and the object. The agent is adapted to receive the state signal, to issue an action capable of generating at least one change in the state of the learning environment, to produce a set of observations relevant to a task, to hypothesize a set of action models configured to explain the observations, and to vet the set of action models to identify a learned model for the task.