Imbuing Human-Robot Teams with Intention
RecognitionDr. Gita [email protected]
Students: Ken Laviers, Bennie Lewis
Funder: DARPA CSSP
Intelligent Agents Lab
SoftwareAgents
Robots(Mechanical Agents)
Humans(Biological Agents)
Intention(Plan,
Activity, Goal)Recognition
Research Problems
Improving plan, activity, and intention recognition Fast Sufficiently accurate Acquiring training data Making it sample-efficient
Determining when to act autonomously Transfer-of-control
Identifying what to do This can be the hardest problem!
Not annoying the human users! Mutual predictability Human must be able to infer the intentions of the agents/robots
Improved game and simulation AI Intelligent training and tutoring systems Human-robot interfaces Elder-care home monitoring systems
Domains
Example: Adaptive Opponents
Exploit adversarial models to improve team decision-making
Create tree of team spatio-temporal traces
Combine output from multiple classifiers to reliably recognize plays
Modify policy of key players to improve play of entire team
Adapt in real-time to the strategy employed by the human player
Online Play Recognition
Learning to Adapt Divide and conquer the
problem into several learning modules Play recognizer Successor state estimator Reward estimator
Individual modules are inaccurate but combine to learn an effective play adaptation.
Use Monte Carlo search to rapidly evaluate large number of play adaptations
Play Recognizer
?
Results
8
Adaptive agents improves the yardage gained in a play and reduce the number of interceptions
over the standard game AI.
Human-Agent-Robot Teams
Robots need humans to use their past experience and common-sense knowledge: To process ambiguous sensor data To solve complicated planning problems (e.g., figuring out the
grasp points on objects) Humans need help with repetitive tasks:
Monitoring multiple information streams (video or audio) Toggling between multiple robots
Agents can facilitate HRI by: Monitoring the humans to identify operator distraction Remembering and propagating commands intelligently across
teams of robots
User Interface
(view from an overhead camera)
User Interface
(gamepad control is popular with our student test subjects)
RSARSim
(video)
Learn Models of User Distraction
Learn model of user distraction by inserting artificial visual distractions into simulation
Identify which of the three robots the user is paying attention to
Features based on robot motion trajectories
Use EM to fit parameters to HMM model
Perform transfer-of control-when distraction levels go over a certain level
Agent
Learn Models of User Distraction
Identification of user distraction level more accurate than models that don’t remember past state
Two state classification accuracy shows our decision threshold (control vs. no-control)
Statistically significant improvements (p<.05) on time required to find the total number of victims in urban rescue scenario
5 class 2 classHMM 37% 88%
SVM 28& 67%
Bayes 28% 67%
DT 28% 67%
Multi-Robot Manipulation Sensors on robot are insufficient
for good grasp planning Toggling rapidly between robots is
complicated for users Idea: leverage commands given
by user to one robot to propagate (and translate) for second robot)
User study evaluating command paradigm: Follow Me: 2nd robot joins the 1st
robot Mirror Me: 2nd robot copies the 1st
robot Scenario involves moving piles of
objects to a goal location, some of which require two robots to move
Human-Agent-Robot Teams
User study on 20 users had very promising results Introducing these two new primitives results in
reductions in both time required to complete the task and in reducing the number of object drops in most of the scenarios
Favorable responses on the post-test questionnaire
Current work: Incorporating a learning by demonstration mode to
allow users to learn the primitives rather than having them preprogrammed by the developer
Conclusions
SoftwareAgents
Robots(Mechanical Agents)
Humans(Biological Agents)
Agents are well-positioned to serve as an enabler of mutual predictability through a combination of intention recognition
and communication monitoring.
Multi-Robot Manipulation