Why HRI needs AI Katia Sycara Robotics Institute Carnegie Mellon - - PowerPoint PPT Presentation

AI and Robotics Interface: Why HRI needs AI

Katia Sycara Robotics Institute Carnegie Mellon University katia@cs.cmu.edu

Why Now?

• Robotic Platforms have improved - robustness
• Commercially available and relatively cheap
• Advances in Vision and Speech Understanding
• Vastly increased computational power
• Networked Capabilities

Result: time to revisit use of AI techniques in Robotics

– Increase autonomous operation of robots – Robots interacting with humans (work, home, play) – Robots in society

Katia Sycara -AI and Robotics 2

Issues in HRI

• Imparting and recognizing intent from human

to robot and vice versa

• Establishing and maintaining “common

ground”

• Effective co-planning, human robot teaming
• Dealing with failures
• Life long learning

Katia Sycara -AI and Robotics 3

Imparting and Recognizing Intent, Plans and Behavior

• Necessary for any complex human in the loop

autonomous robotic system

• Two way street –especially challenging in

multi-robot systems (AI has dealt with those)

– Robot recognizes human intent – Human recognizes robot intent and behavior – Plan recognition and monitoring – Failure detection and repair – Robot generates explanations of behavior

Katia Sycara -AI and Robotics 4

Problems of common ground: Perception/ common referents

What a human sees What the robot sees A man with a walker relatively close to observer An obstacle at distance of 183 cm at 15⁰

Open Problems

• Signal to symbol mapping
• Representing context

Katia Sycara -AI and Robotics 5

Characteristics of Robot Assistance and Teaming

• Unobtrusiveness: the robot should assist the human without becoming a nuisance
• Timeliness: the robot will not introduce delays to human or to itself
• Non-interruption: robot will not interrupt human inappropriately
• Safety: the robot should mind human’s and its own safety
• Predictability: human should be able to interpret and predict robot’s actions; robot

should be able to recognize and predict humans actions

• Fluency: the robot should assist/interact with human for appropriate tasks at

appropriate time

• Proactivity: the robot should perform appropriate assistive actions without having

been asked

• Goal achievement: robot should know when the human-robot system has

achieved the goal

• Robustness/failure recovery: the robot should be able to gracefully recover from

failure and perform repairs in a way that preserves interaction coherence

• Need for formalization and computational modeling of the above characteristics,
• Need for performance metrics

These are components of trusted interaction

Katia Sycara -AI and Robotics 6

AIR Issues in Human Robot Collaboration

• Representation:

– Need representations that span high level tasks to low level robotic motions, semantic representation of the environment (function, physical qualities etc) – Discrete to continuous – Signal to Symbol (eg semantic labeling of images) – Abstraction levels – Represent human actions, human and robot capabilities so that good functional allocation can be done or emerge

• Inference

– Computational tractability of symbolic representation schemes (dimensionality of space of potential robot actions even larger) – Scalable algorithms for reasoning about uncertainty and dynamism – Recognizing and dealing with failure – Common sense reasoning

Katia Sycara -AI and Robotics 7

AI-R Issues in Human Robot Collaboration

• Learning from experience

– What experiences to collect and how, so they are reusable across different contexts – Representation, storage, indexing, retrieval of past experiences – Data mining not enough- need for semantics, data to knowledge – Experiences with high level tasks should be coherent with low level robot planning – Lifelong Learning – Learning from Failure

Katia Sycara -AI and Robotics 8

Steps towards addressing the issues

• Some good progress has been made in combinations of

logical and probabilistic representations (e.g. probabilistic HTNs)

• Progress in hybrid systems, verification techniques

extending to MAS

• Web accessible ontologies and representation schemes for

high level task knowledge and low level motion planning, path planning etc

• Human-based computation to provide data for learning by

demonstration

• Computational intractability for complex tasks will not go

away: hence discover and exploit structure of different tasks and devise good approximation algorithms

Katia Sycara -AI and Robotics 9

Science and Engineering

• Science-Simplifying assumptions so that

formal models can be constructed

• Engineering-heuristics, dealing with

complexity, integration of components How to best synergize these?

Katia Sycara -AI and Robotics 10

• Questions?

Katia Sycara -AI and Robotics 11