TASC: Teammate Algorithm for Shared Cooperation

Link to paper

Conference: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020

Authors: Mai Lee Chang, Taylor Kessler Faulkner, Thomas Benjamin Wei, Elaine Schaertl Short, Gokul Anandaraman, Andrea Lockerd Thomaz

My Role: Lead Researcher

Research Overview

This research introduces TASC (Teammate Algorithm for Shared Cooperation), a novel algorithm that enables robots to function as effective teammates to humans. TASC is grounded in Bratman's concept of shared cooperative activity (SCA), which defines three essential facets of teamwork:

1. Mutual Responsiveness (MR): Appropriately reacting to the intentions and actions of teammates

2. Commitment to the Joint Activity (CJA): Aligning sub-plans so all team members participate in the same activity

3. Commitment to Mutual Support (CMS): Willingness to help teammates if there are breakdowns

Innovation

TASC transforms these sociological concepts into a computational framework with three weighted parameters:

• Legibility: Enables mutual responsiveness by recognizing and communicating intent

• Effort: Demonstrates commitment to the joint activity by taking actions that appear effortful

• Value: Shows commitment to mutual support by providing assistance toward achieving the team's goal

Technical Implementation

The algorithm operates within a Markov Decision Process (MDP) framework where:

• The robot predicts the human's goal using a classifier

• The robot takes legible actions to convey its intended goal

• The robot takes actions that appear effortful to the human

• The robot selects actions that maximize the expected value toward completing the task

The robot's action selection is based on the weighted sum of these three parameters, which can be tuned according to the specific requirements of different tasks.

Experimental Validation

We conducted three user studies on Amazon Mechanical Turk to evaluate TASC:

Study 1: Navigation Task

• Task: Human and robot worked together to move a remote-controlled car to goal states

• Experimental design: One-way between-subjects (independent variable is weight setting)

• Participants: 153 total, 51 in each condition (64 F, 85 M, and 4 preferred not to answer, age: Mean = 35.07, SD = 10.59)

• Findings: Prioritizing value or weighting all parameters equally resulted in significantly better performance compared to prioritizing legibility

Study 2: Modified Navigation Task

• Task: Robot knew the goal, but the human did not

• Experimental design: One-way within-subjects (independent variable is the weight setting)

• Participants: 34 (17 M, 17 F, age: Mean = 34.50; SD = 9.98)

• Findings: Giving weight to legibility allowed participants to make accurate goal predictions earlier (by one robot move) and with more confidence

Study 3: Tower Assembly Task

• Task: Human and robot collaboratively built towers using blocks

• Experimental design: One-way within-subjects (independent variable is the weight setting)

• Participants: 28 (10 F, 17 M and 1 preferred not to answer, age: Mean = 29.14; SD = 6.00)

• Findings: Giving weight to effort enabled participants to use 6% less energy and positively influenced their perception of mutual responsiveness

Key Findings

Our research on the TASC algorithm revealed three key insights:

1. Task-Specific Parameter Tuning: Different collaborative tasks require different weightings of legibility, effort, and value parameters - prioritizing value worked best for efficiency-focused tasks, while balancing all parameters improved overall teamwork quality.

2. Enhanced Human Understanding: Prioritizing legibility allowed humans to predict the robot's goals one move earlier and with significantly greater confidence (76% of participants preferred this condition), without compromising task efficiency.

3. Improved Team Dynamics: When the robot demonstrated appropriate effort, humans used 6% less energy and rated mutual responsiveness significantly higher, indicating that perceived robot commitment positively influences human behavior and perception of teamwork.

Research Impact

This research makes several notable contributions:

• Translates a sociological framework of human-human teamwork into a computational model for human-robot interaction

• Demonstrates that considering all three facets of shared cooperative activity improves perceived teamwork

• Shows how different tasks require different weightings of legibility, effort, and value

• Creates a foundation for adaptive teamwork algorithms that can adjust to various collaborative scenarios

Future Directions

This work opens up a multi-dimensional space for teamwork that can be explored through:

• Reinforcement learning to find optimal weight combinations for different tasks

• Extension to physical robot implementations

• Application to more complex collaborative scenarios

• Integration with intent recognition systems

Skills Demonstrated

Technical Skills

• Algorithm Design: Created a novel algorithm for robot action selection in collaborative settings

• Markov Decision Processes: Implemented and solved MDPs for sequential decision-making

• Probabilistic Reasoning: Used probabilistic models for goal and action prediction

• Robot Intent Communication: Designed systems for making robot intentions legible to humans

• Software Development: Implemented simulations for human-robot interaction experiments (Python, MDPtoolbox, HTML)

Research Skills

• Experimental Design: Designed three user studies to evaluate the algorithm

• Human-Subject Research: Conducted studies with human participants on Amazon Mechanical Turk

• Data Analysis: Performed statistical analysis of objective and subjective measures using R and Python (Cronbach’s alpha, ANOVA, t-tests, Likert surveys)

• Literature Synthesis: Incorporated concepts from philosophy and cognitive science into robotics

• Scientific Communication: Presented complex technical concepts clearly in academic writing and conference presentation

Domain Knowledge

• Human-Robot Interaction: Deep understanding of how humans and robots can work together

• Computational Models of Teamwork: Expertise in formalizing social concepts computationally

• Robotics: Knowledge of robot planning, perception, and action execution

• Human Factors: Understanding of how humans perceive and interpret robot behavior

• Collaborative Systems: Experience designing systems where humans and AI collaborate