Method and system for machine concept understanding
AssigneeIntrinsic Innovation LLC
InventorsMiguel Lázaro-Gredilla, Dianhuan Lin, J. Swaroop Guntupalli, Dileep George
Life of this patent
Estimating the term…
Abstract
A system and method for machine understanding, using program induction, includes a visual cognitive computer including a set of components designed to execute predetermined primitive functions. The method includes determining programs using a program induction engine that interfaces with the visual cognitive computer to discover programs using the predetermined primitive functions and/or executes the discovered programs based on an input.
Claims22
A method for vision-based program induction, comprising:
- receiving at least one training set, the training set consisting essentially of an input scene and a target scene;
- detecting an object within the input scene and the target scene with a generative vision model;
- automatically selecting a series of primitive functions from a candidate set of primitive functions based on the input scene and the target scene, the series of primitive functions comprising:
- robotic primitive functions;
- and
- vision hierarchy primitive functions for the generative vision model;
- simulating a series of object positions within a simulated scene based on the series of primitive functions and a set of argument values, with a dynamics model and the generative vision model;
- and
- selecting the series of primitive functions and the set of argument values as the program in response to a final simulated object position substantially matching a target position of the object in the target scene, wherein the program is executed using a new input scene without a target inference scene.
A system for robotic program execution, comprising at least one processor configured to execute:
- a vision hierarchy that detects objects within an input scene and extracts a set of attributes for each object from an input scene, the vision hierarchy responsive to a set of vision hierarchy primitive functions;
- a fixation controller that positions a reference point of a field of interest within the input scene, the fixation controller responsive to a set of fixation control primitive functions;
- an attention controller that selects an attended set of objects within the field of interest from the detected objects, the attention controller responsive to a set of attention control primitive functions;
- a robot controller that controls a manipulator to manipulate the selected attended set of objects within the input scene, the robot controller responsive to a set of robot control primitive functions;
- and
- a program comprising a series of primitive functions that are selected based on moving an object in an input scene to a final simulated object position that substantially matches a target position of the object in a target scene, the program comprising:
- vision hierarchy primitive functions from the set of vision hierarchy primitive functions;
- fixation control primitive functions from the set of fixation control primitive functions;
- attention control primitive functions from the set of attention control primitive functions;
- and robot control primitive functions from the set of robot control primitive functions, wherein the program is executed using a new input scene without a target inference scene.
Dependencies are inferred from the claim text.