Understanding Robot Minds: Leveraging Machine Teaching for Transparent Human-Robot Collaboration Across Diverse Groups

TL;DR

This work addresses transparent human-robot collaboration in groups with diverse learning abilities by developing team-belief machine-teaching methods. It combines a particle-filter representation of individual learner beliefs with two team-belief models (common and joint) and a closed-loop teaching curriculum that samples informative demonstrations via counterfactuals. Across simulations, group-belief strategies generally outperform individual strategies in terms of learning efficiency, with joint belief providing the most informative demonstrations for proficient-dominated teams and individual strategies excelling for homogeneous-naive groups; real-time proficiency adaptation emerges as key for optimal teaching. The study highlights the importance of adaptive, group-level explainability in mixed-proficiency settings and lays groundwork for robust, real-time team-based robot policy communication in diverse environments.

Abstract

In this work, we aim to improve transparency and efficacy in human-robot collaboration by developing machine teaching algorithms suitable for groups with varied learning capabilities. While previous approaches focused on tailored approaches for teaching individuals, our method teaches teams with various compositions of diverse learners using team belief representations to address personalization challenges within groups. We investigate various group teaching strategies, such as focusing on individual beliefs or the group's collective beliefs, and assess their impact on learning robot policies for different team compositions. Our findings reveal that team belief strategies yield less variation in learning duration and better accommodate diverse teams compared to individual belief strategies, suggesting their suitability in mixed-proficiency settings with limited resources. Conversely, individual belief strategies provide a more uniform knowledge level, particularly effective for homogeneously inexperienced groups. Our study indicates that the teaching strategy's efficacy is significantly influenced by team composition and learner proficiency, highlighting the importance of real-time assessment of learner proficiency and adapting teaching approaches based on learner proficiency for optimal teaching outcomes.

Figures (6)

• Figure 1: The figure illustrates the complexity of group machine teaching, highlighting the disparity in understanding from common examples among diverse group members. Personalizing examples to a group is challenging due to varied individual beliefs and learning abilities. Our approach utilizes estimations of individual and collective team beliefs to tailor demonstrations for effective communication of the robot's policy to the entire group.
• Figure 2: Update process of a learner's belief represented by a particle filter. A cross-section of the custom probability density function (pdf) used to update particle weights is shown. Particles consistent with the demonstrated behavior receive higher weights via a uniform distribution (yellow ring), while those on the inconsistent side are weighted less, decreasing exponentially with distance from the constraint, via a von-Mises Fisher distribution. The updated belief distribution is shown on the right.
• Figure 3: This figure illustrates an example set of test responses for a team with two individuals, P1 and P2. The test responses are transformed to constraints. The yellow partial spheres show the regions that are consistent with their test response, i.e. agree with the constraint. When their responses are different, the constraints space of their common belief of their tests is their intersection of individual beliefs and that of their joint belief is the union of individual beliefs as depicted. These constraints spaces are used to update the weights of the PF distributions.
• Figure 4: Interactions and corresponding PF belief updates for a team with three people for the first interaction period. The red particle represents the true reward weight. An interaction period consists of one set of demos related to a KC, followed by a set of tests, and then feedback (corrective or confirmatory). After the demos, all individual and team beliefs are updated based on expected information gain from demos. The distributions are similar since all individuals are expected to learn equally. After the demos, they are provided with test(s) to evaluate their understanding and their responses are used directly for updating individual beliefs and aggregated to update team beliefs. In this case, P1 and P3 got the response correctly and P2 got it incorrect. The difference in the constraint spaces and the updated distributions after tests can be observed for the individual and team beliefs. Confirmatory or corrective feedback is given after the tests and they are expected to learn from either feedback. The distributions are updated to reflect this learning from feedback.
• Figure 5: Experimental results on the effects of demonstration strategy and team composition. (a) All group teaching strategies performed better than the baseline strategy of teaching individuals sequentially in terms of number of interactions. No discernable difference due to strategies for number of interactions. Expected differences in teams, teams with more proficient learners learned quicker, observed. (b) Noticeable differences in average team knowledge observed for strategy but differences are not statistically significant. Also observed that teams with more proficient learners had higher knowledge level.