Interpreting human-robot instructions
A system includes a robot having a module that includes a function for mapping natural language commands of varying complexities to reward functions at different levels of abstraction within a hierarchical planning framework, the function including using a deep neural network language model that learns how to map the natural language commands to reward functions at an appropriate level of the hierarchical planning framework.
1. A system comprising:
a robot comprising a module to interpret natural language commands, the module comprising a function for mapping natural language commands of varying complexities to reward functions at different levels of abstraction of states and actions within a hierarchical planning framework, wherein the hierarchical planning framework comprises an Abstract Markov Decision process, the function comprising:
a deep neural network language model that learns how to map the natural language commands to reward functions at an appropriate level of the hierarchical planning framework to perform navigation and object manipulation tasks using the robot, the Abstract Markov Decision process comprising:
a set of states that define an environment specified in an object-oriented fashion with object classes and attributes;
a set of actions an agent can execute to transition between states or to invoke a lower-level AMDP subtask;
a transition probability distribution over all possible next states given a current state and executed action;
a numerical reward earned for a particular transition;
a discount factor or effective time horizon under consideration; and
a state projection function that maps lower-level states to higher-level (AMDP) states.
2. A system comprising:
a robot comprising a module comprising a function for mapping natural language commands of varying complexities to reward functions at different levels of abstraction of states and actions within a hierarchical planning framework to perform navigation and object manipulation tasks using the robot, wherein the hierarchical planning framework comprises an Abstract Markov Decision process, the function using a deep neural network language model that learns how to map the natural language commands to reward functions at an appropriate level of the hierarchical planning framework, the function further comprising an Abstract Markov Decision Process to represent a decision-making problem involving a hierarchy of the robot's states and actions, the Abstract Markov Decision Process comprising:
a set of states that define an environment specified in an object-oriented fashion with object classes and attributes;
a set of actions an agent can execute to transition between states or to invoke a lower-level AMDP subtask;
a transition probability distribution over all possible next states given a current state and executed action;
a numerical reward earned for a particular transition;
a discount factor or effective time horizon under consideration; and
a state projection function that maps lower-level states to higher-level (AMDP) states.
3. A method comprising:
providing a mobile-manipulator robot; and
providing a module comprising a function that interprets and grounds natural language commands to a mobile-manipulator robot at multiple levels of abstraction of states and actions, the function comprising a deep neural network language model that learns how to map the natural language commands to reward functions at an appropriate level of a hierarchical planning framework to perform navigation and object manipulation tasks using the mobile-manipulator robot, the function further comprising an Abstract Markov Decision process to represent a decision-making problem involving a hierarchy of the robot's states and actions, the Abstract Markov Decision process comprising:
a set of states that define an environment specified in an object-oriented fashion with object classes and attributes;
a set of actions an agent can execute to transition between states or to invoke a lower-level AMDP subtask;
a transition probability distribution over all possible next states given a current state and executed action;
a numerical reward earned for a particular transition;
a discount factor or effective time horizon under consideration; and
a state projection function that maps lower-level states to higher-level (AMDP) states.