Robotics: Science and Systems, July 12-16, 2020 Iterative Repair of Social Robot Programs from Implicit User Feedback via Bayesian Inference Michael Jae-Yoon Chung and Maya Cakmak
 University of Washington, Computer Science & Engineering

Photo: USC Robotics Research Lab Photo: MIT Media Lab/Personal Robots Group Photo: Yale Social Robotics Lab

Photo: USC Robotics Research Lab Photo: Yale Social Robotics Lab Photo: MIT Media Lab/Personal Robots Group Photo: MIT Media Lab/Personal Robots Group Photo: MIT Media Lab/Personal Robots Group Photo: MIT Media Lab/Personal Robots Group

Research Goal Facilitate fluent and autonomous social robot behavior authoring with an expressive yet maintainable approach

No content

program sketch 1 lookingAtRobot = debounce(abs(faceYawAngle < ??minDisengagedAngle), ??disengagedTimeout); ... if (curState == "read" && sayFinished && humanAvailable && !allDone) { return "read"; } else if (curState == "read" && (sayFinished && humanBusy && !allDone || goBackTapped)) { return "wait"; } else if (curState == "read" && sayFinished && allDone) { return "stop"; } else if (curState == "wait" && (humanAvailable || nextTapped && !hold))) { return "read"; } Hole variable name Distribution ??minDisengagedAngle Categorical ??disengagedTimeout Categorical ... ...

Interaction trace program sketch 1 interaction0 2 Note: the robot in the video is not running the example program

program sketch 1 interaction0 2 interaction trace 3 correction0 4 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700 1800 1900 500 say wait read wait goBackTapped Time (ms) inputs outputs state trace wait corrected state trace lookingAtRobot Error Correction

600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700 1800 1900 500 say wait wait read wait goBackTapped Time (ms) inputs outputs state trace corrected state trace lookingAtRobot Error Correction User feedback program sketch 1 interaction0 2 interaction trace 3 correction0 4

program sketch 1 interaction0 2 interaction trace 3 correction0 4 (repaired) Interaction0 5 ... ... ... ... ... ... At each iteration, do 1. Run the program and collect an interaction trace 2. Obtain corrections from the programmer or the user 3. Re-estimate hole distributions using the input and corrected state traces (posterior) 4. Fill-in holes using maximum a posteriori estimation (MAP) for the next iteration

• Human simulators for sampling input and ground truth state trace • Baseline: search-based repair • 3 scenarios, 5 repair iterations • Measures • % overlap (interaction quality) • Repair time in ms (speed) A top-down human-simulator. Samples from states, substates (intent), and then inputs Simulation Experiment: Setup

0 1 2 3 4 5 repair iterations 0.0 0.2 0.4 0.6 0.8 1.0 1.2 overlap % Storytelling 0 1 2 3 4 5 repair iterations 0.0 0.2 0.4 0.6 0.8 1.0 1.2 Neck Exercise 0 1 2 3 4 5 repair iterations 0.0 0.2 0.4 0.6 0.8 1.0 1.2 Open-ended Q&A bayesian search method Simulation Experiment: % Overlap Results

0 1 2 3 4 5 repair iterations 0.0 0.2 0.4 0.6 0.8 1.0 1.2 overlap % Storytelling 0 1 2 3 4 5 repair iterations 0.0 0.2 0.4 0.6 0.8 1.0 1.2 Neck Exercise 0 1 2 3 4 5 repair iterations 0.0 0.2 0.4 0.6 0.8 1.0 1.2 Open-ended Q&A bayesian search method Simulation Experiment: Speed Results 1 2 3 4 5 repair iterations 0 2,000 4,000 6,000 8,000 10,000 12,000 time (ms) Storytelling 1 2 3 4 5 repair iterations 0 1,000 2,000 3,000 4,000 Neck Exercise 1 2 3 4 5 repair iterations 0 20,000 40,000 60,000 80,000 100,000 120,000 Open-ended Q&A bayesian search method

• 10 participants • 7 males, 23.2 yrs old on average • A custom-built social robot • 2 scenarios, 4 iterations Human experiment setup. Human Experiment: Setup

Qualitative Results: Before Repair

Qualitative Results: After Repair

More Results and Discussions • Simulation experiments • Noisy corrections study • Human experiments • Quantitative results • Findings from data analysis • Robot behavior adaptation experiment • Discussions and future direction Thank you!