1 / 15

Behavior Recognition and Opponent Modeling in Autonomous Multi-Robot Systems

Behavior Recognition and Opponent Modeling in Autonomous Multi-Robot Systems. Keith J. O’Hara College of Computing Georgia Institute of Technology kjohara@cc.gatech.edu. Introduction. Recognizing and modeling behavior from low-level action thru high-level strategy.

enye
Download Presentation

Behavior Recognition and Opponent Modeling in Autonomous Multi-Robot Systems

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Behavior Recognition and Opponent Modeling inAutonomous Multi-Robot Systems Keith J. O’Hara College of Computing Georgia Institute of Technology kjohara@cc.gatech.edu

  2. Introduction • Recognizing and modeling behavior from low-level action thru high-level strategy. • Single agent primitive action • A sequence of single agent actions • Group behavior • To understand opponents • To understand teammates • No Communication • Communication troublesome or dangerous • Speak different “languages” • Operate based on a different behavior vocabulary

  3. Outline • 2 Approaches • Intille and Bobick (MIT) • Application of bayesian belief networks for American football play recognition. • Han and Veloso (CMU) • Behavior Hidden Markov Models for robot soccer behavior recognition.

  4. Important Themes • Single/Multi agent • Recognition of agents and primitive actions • Agent subgoals, goals, intentions • Group subgoals, goals, intentions • Online recognition • Uncertainty in Perception • Uncertainty/Flexibility of Plan • Use of probabilistic techniques to deal with uncertainty. • Completely described action and observation spaces.

  5. “Recognizing Multi-Agent Action from Visual Evidence” • Recognition of American football plays from real games. • Assumes we have labeled participants with rough position and orientation estimates. • Properties of the domain: • Complex: partially ordered causal events • Multi-agent: parallel event streams • Uncertain: Uncertainty in both data and model • Other domains • Sports, military, traffic, robotics

  6. Method • Method inspired by model-based object recognition techniques. • Database of plays (temporal structure descriptions) described by temporal and logical relationships of events. • Construct “visual network” to detect individual goals (primitive actions) from visual evidence.

  7. Temporal Structure Descriptions • Individual Goal • Action Components • Object Assignment • Temporal Constraints

  8. Visual Networks • Construct belief network (visual network) based upon visual evidence.

  9. Multi-Agent Belief Network • Multi-Agent Networks normally contain at least 50 belief nodes and 40 evidence nodes • Conditional and prior probabilities are determined automatically

  10. Results • System of 29 tracked plays, 10 temporal play descriptions • 21/25 were recognized correctly • False positives are a problems. (plays that aren’t defined) • Recognized single-agent behavior and multi-agent plays. • Handled fuzzy temporal relationships (around, before). • Not evaluated online. • Assumes tracking/labeling/localization problem is solved. (Manually done in this work.) • Must know entire domain of observations (player states), and all possible plans (play book).

  11. “Automated Robot Behavior Recognition” • Robot Soccer • Adaptable Strategy • Narrative Agents • Coaches • Formalism • Agent R is the observed robot • Agent O is the observing robot • R acts according to a known set of behaviors h(i) • O has a model of the set of the possible behaviors. • O must decide which h(i), R is performing. • Must be online algorithm. • One observed robot and one observed ball.

  12. Go-To-Ball s1 s2 s3 O1 O2, O3 O3 s4 O1 O2 O1 O3 Method(1) • Use Hidden Markov Models (HMMs) to recognize behaviors • Motivated by success of HMMs in other “recognition” tasks. (e.g. speech, gesture) • A Behavioral HMM() for each behavior • Set of States • Initial, intermediate, accept, reject • Observations Space • Absolute/Relative Position, Dynamic (velocity) • State Transition Matrix • Observation Probabilities • Initial State Distribution • P(this state | observations, )

  13. Go-To-Ball s1 s2 s3 O1 O2, O3 O3 s4 O1 O2 O1 O3 Method(2) • The BHMM() • Set of States • Observations Space • State Transition Matrix • Observation Probabilities • Initial State Distribution

  14. Results • Online algorithm • Applied to robotics domain (simulation/real-robots) • Implemented everyone’s favorite behaviors • Go-To-Ball, Go-Behind-Ball, Intercept-Ball, Goalie-Align-Ball • Not much quantitative evidence. • Only single agent case. • Assume each behavior to be a sequence of state traversals. • BHMM and behavior initial states must match up, or use a timeout/restart mechanism. • Mentioned by Intille and Bobick as a problem with treating temporal constraints as first-order markovian.

  15. Conclusions • New and hard problem. • Use of probabilistic techniques to deal with uncertainty in perception and the plan. • Completely described action and observation spaces.

More Related