Systems and methods for robotic teleoperation intention estimation
Systems and methods for robotic teleoperation are provided. The system receives motion input and gaze data of a human operator performing a robotic teleoperation task, and a plurality of features are extracted from the received data. A recurrent neural network determines latent actions from the extracted features and generates temporally associated action inferences. A model determines inferred intention from a sequenced pair of action inferences, an intention hierarchy, and a probabilistic uncertainty function. A teleoperated robotic element performs a task associated with the inferred intention without a corresponding control signal from the human operator. The recurrent neural network is trained using extracted features and contrastive learning. The intention hierarchy is generated from action inferences. The probabilistic uncertainty function is generated using Bayesian hierarchical modeling. The system and method improves robotic teleoperation by estimating the intention of the human operator.
1 . A system for robotic teleoperation, comprising:
a robotic teleoperation input configured to:
receive a plurality of motion inputs of a human operator while performing a robotic teleoperation task, and
receive a plurality of gaze data of the human operator while performing the robotic teleoperation task;
a robotic teleoperation feature extractor configured to:
analyze the motion inputs and the gaze data associatively, and
extract a plurality of features from the motion inputs and the gaze data;
a robotic teleoperation neural network configured to:
determine a plurality of latent actions, each associated with one or more extracted features, and
generate a plurality of temporally associated action inferences, each action inference being associated with a determined latent action and associated extracted features;
a robotic teleoperation transition model configured to:
determine elements of a transition matrix associated with a sequenced pair of action inferences;
a robotic teleoperation dynamic transition model configured to:
determine an inferred intention associated with the determined elements of the transition matrix,
wherein the dynamic transition model includes an intention hierarchy associated with a plurality of inferred intentions, and
wherein the inferred intention is based at least in part on an estimated probabilistic uncertainty applied to the intention hierarchy; and
a teleoperated robotic element that performs a task associated with the inferred intention without a corresponding control signal from the human operator.
2 . The system of claim 1 , wherein the robotic teleoperation neural network is further configured to train a recurrent neural network with the extracted features.
3 . The system of claim 1 , wherein the robotic teleoperation transition model generates one or more transition matrices associated with one or more sequenced pairs of action inferences.
4 . The system of claim 3 , wherein the robotic teleoperation transition model is refined using contrastive learning when a sequenced pair of action inferences is not associated with an intended teleoperation of a robotic element.
5 . The system of claim 3 , wherein the robotic teleoperation transition model performs a contrastive loss function to generate weights associated with one or more transition matrices and applies a conditional rule to a transition matrix to redetermine at least one action inference associated with the transition matrix.
6 . The system of claim 1 , wherein the robotic teleoperation dynamic transition model is further configured to:
perform hierarchy discovery using a hierarchical clustering algorithm to generate an intention hierarchy from action inferences;
estimate probabilistic uncertainty of at least a part of the intention hierarchy uncertainty using Bayesian hierarchical modeling; and
generate a likelihood estimation of an inferred intention based on the estimated probabilistic uncertainty applied to the intention hierarchy.
7 . The system of claim 6 , wherein the hierarchical clustering algorithm includes a Ward's linkage algorithm, and wherein Bayesian hierarchical modeling includes one or more of a Markov Chain Monte Carlo and a Variational Inference algorithm.
8 . A method for robotic teleoperation, comprising:
receiving a plurality of gaze data and a plurality of motion inputs associated with a human operator performing a robotic teleoperation task;
extracting a plurality of features from the motion inputs and the gaze data;
determining, by a neural network, a plurality of latent actions, each associated with one or more extracted features;
generating, by the neural network, a plurality of temporally associated action inferences, each action inference being associated with a determined latent action and associated extracted features;
determining elements of a transition matrix associated with a sequenced pair of action inferences;
determining, using a dynamic transition model, an inferred intention associated with the determined elements of the transition matrix,
wherein the dynamic transition model includes an intention hierarchy associated with a plurality of inferred intentions, and
wherein the inferred intention is based at least in part on an estimated probabilistic uncertainty applied to the intention hierarchy; and
performing, by a teleoperated robotic element, a task associated with the inferred intention without a corresponding control signal from the human operator.
9 . The method of claim 8 , wherein neural network is a recurrent neural network and further comprising:
training the recurrent neural network with the extracted features.
10 . The method of claim 8 , further comprising:
generating one or more transition matrices associated with one or more sequenced pairs of action inferences.
11 . The method of claim 10 , further comprising:
refining the robotic teleoperation transition model using contrastive learning when a sequenced pair of action inferences is not associated with an intended teleoperation of a robotic element.
12 . The method of claim 10 , further comprising:
performing a contrastive loss function to generate weights associated with one or more transition matrices; and
applying a conditional rule to a transition matrix to redetermine at least one action inference associated with the transition matrix.
13 . The method of claim 8 , further comprising:
performing hierarchy discovery using a hierarchical clustering algorithm to generate an intention hierarchy from action inferences;
estimating probabilistic uncertainty of at least a part of the intention hierarchy uncertainty using Bayesian hierarchical modeling; and
generating a likelihood estimation of an inferred intention based on the estimated probabilistic uncertainty applied to the intention hierarchy.
14 . The method of claim 13 , wherein the hierarchical clustering algorithm includes a Ward's linkage algorithm, and wherein Bayesian hierarchical modeling includes one or more of a Markov Chain Monte Carlo and a Variational Inference algorithm.
15 . A non-transitory computer readable storage medium storing instructions that when executed by a computer having a processor to perform a method for robotic teleoperation, the method comprising:
receiving a plurality of gaze data and a plurality of motion inputs associated with a human operator performing a robotic teleoperation task;
extracting a plurality of features from the motion inputs and the gaze data;
determining, by a neural network, a plurality of latent actions, each associated with one or more extracted features;
generating, by the neural network, a plurality of temporally associated action inferences, each action inference being associated with a determined latent action and associated extracted features;
determining elements of a transition matrix associated with a sequenced pair of action inferences;
determining, using a dynamic transition model, an inferred intention associated with the determined elements of the transition matrix,
wherein the dynamic transition model includes an intention hierarchy associated with a plurality of inferred intentions, and
wherein the inferred intention is based at least in part on an estimated probabilistic uncertainty applied to the intention hierarchy; and
performing, by a teleoperated robotic element, a task associated with the inferred intention without a corresponding control signal from the human operator.
16 . The non-transitory computer readable storage medium of claim 15 , wherein neural network is a recurrent neural network and further comprising:
training the recurrent neural network with the extracted features.
17 . The non-transitory computer readable storage medium of claim 15 , further comprising:
generating one or more transition matrices associated with one or more sequenced pairs of action inferences.
18 . The non-transitory computer readable storage medium of claim 17 , further comprising:
refining the robotic teleoperation transition model using contrastive learning when a sequenced pair of action inferences is not associated with an intended teleoperation of a robotic element.
19 . The non-transitory computer readable storage medium of claim 17 , further comprising:
performing a contrastive loss function to generate weights associated with one or more transition matrices; and
applying a conditional rule to a transition matrix to redetermine at least one action inference associated with the transition matrix.
20 . The non-transitory computer readable storage medium of claim 15 , further comprising:
performing hierarchy discovery using a hierarchical clustering algorithm to generate an intention hierarchy from action inferences;
estimating probabilistic uncertainty of at least a part of the intention hierarchy uncertainty using Bayesian hierarchical modeling; and
generating a likelihood estimation of an inferred intention based on the estimated probabilistic uncertainty applied to the intention hierarchy.