Mixed-Initiative Human-Robot Teaming under Suboptimality with Online Bayesian Adaptation

TL;DR

This work develops computational modeling and optimization techniques for enhancing the performance of human-agent teams, where both the human and the robotic agent have asymmetric capabilities and act suboptimally due to incomplete environmental knowledge.

Abstract

For effective human-agent teaming, robots and other artificial intelligence (AI) agents must infer their human partner's abilities and behavioral response patterns and adapt accordingly. Most prior works make the unrealistic assumption that one or more teammates can act near-optimally. In real-world collaboration, humans and autonomous agents can be suboptimal, especially when each only has partial domain knowledge. In this work, we develop computational modeling and optimization techniques for enhancing the performance of suboptimal human-agent teams, where the human and the agent have asymmetric capabilities and act suboptimally due to incomplete environmental knowledge. We adopt an online Bayesian approach that enables a robot to infer people's willingness to comply with its assistance in a sequential decision-making game. Our user studies show that user preferences and team performance indeed vary with robot intervention styles, and our approach for mixed-initiative collaborations enhances objective team performance ($p<.001$) and subjective measures, such as user's trust ($p<.001$) and perceived likeability of the robot ($p<.001$).

Figures (5)

• Figure 1: Graphical overview of the Bayes-POMCP approach for mixed-initiative Human-Robot Teaming: At each timestep $t$, the human first takes an action based on interaction history, $h$, and their current observation of the world state, $x$. The robot then determines when and how to intervene by anticipating human behavior using a Monte-Carlo tree search. The reward is calculated based on both human and robot actions.
• Figure 2: Frozen Lake Domain used in this study. Figure 2(a) shows the overall game layout. Figure 2(b) depicts robot intervention styles: interrupt, take-control, and Figure 2(c) shows the human and robot accuracies in identifying slippery grids.
• Figure 3: Results from Data Collection Study with Heuristic Mixed-Initiative Policies. Figure \ref{['fig:score_1']} shows that the team performance is the highest for the take-control agents and the lowest with no-assist (baseline). Figure \ref{['fig:pref_1']} shows the users preference ranking across intervention styles. The majority of the users prefer to work with the interrupt+explain agent the most (rank $= 5$).
• Figure 4: Team performance in simulation experiments with static and dynamic latent user models. Figures \ref{['fig:static_expertise']} and \ref{['fig:static_trust']} show that Bayes-POMCP can enhance team performance across users of varied expertise and compliance tendencies, respectively. Bayes-POMCP outperforms heuristics and the ablation POMCP model, especially for users with low expertise.
• Figure 5: Results from the Evaluation Study. Figure \ref{['fig:score_2']} shows that the team performance is the highest for the Bayes-POMCP agent and the lowest for the Adv-Bayes-POMCP (the adversarial baseline). Figure \ref{['fig:pref_2']} shows that the majority of the users prefer our approach compared to the baselines.