Sequence-to-sequence language grounding of non-Markovian task specifications
A method includes enabling a robot to learn a mapping between English language commands and Linear Temporal Logic (LTL) expressions, wherein neural sequence-to-sequence learning models are employed to infer a LTL sequence corresponding to a given natural language command.
1. A method comprising:
enabling a robot to learn a mapping between English language commands and Linear Temporal Logic (LTL) expressions, wherein neural sequence-to-sequence learning models are employed to infer a LTL sequence corresponding to a given natural language command, wherein geometric LTL (GLTL) paired with a Markov Decision Process (MDP) is used to find policies corresponding to LTL expressions.
2. The method of claim 1 wherein the Markov Decision Process is a five-tuple (S, A, R, T, γ) where S defines the robot state space, A specifies the actions available to the robot, R encodes the underlying task through numerical rewards defined for each state-action pair, T defines the probabilistic transition dynamics of the environment, and γ is a discount factor.
3. The method of claim 2 wherein given the Markov Decision Process as input, a planning algorithm produces a policy that maps from states to robot actions.
4. The method of claim 3 wherein each atomic proposition of a GLTL expression has an associated three-state MDP consisting of an initial, accepting, and rejecting state.