Explaining to and being explained by a service robot: Four HRI studies revisited under a framework for explainability

Robotic butlers relieving humans of domestic chores are currently highly researched and debated, envisioning a near future where these systems will smoothly navigate our private spaces and interact with us naturally and transparently. This vision will hardly be realized as long as potential users are not able to interpret and trust such autonomous systems. This requires negotiating a common interpretation of the environment,  the service to be provided,  and the means of communication.  

While developing a robotic system capable of learning from users and interacting in a supportive manner, we iteratively explored and assessed the technical and communication requirements for various human-robot interaction scenarios. Here, we sum up results from four studies, revisiting them from the perspective of the conceptual framework proposed in Rohlfing et al.  and computationally formalized in.  This framework postulates that explanations are cooperatively constructed through monitoring by the explainer of the information needs of the explainee and through scaffolding, i.e., the undertaking of specific actions by the explainer to facilitate the explainee’s understanding. Robots and humans have different perception and reasoning capabilities, and such a discrepancy can be a cause of misalignment. Further, world knowledge cannot be assumed as common ground; thus, every interaction with a new robot entails bridging multiple explicit and implicit epistemic gaps. Often, humans need to figure out what the robot can see, do, or understand, while the robot needs to figure out what the user wants and translate it in its own representation. So, explaining cues are exchanged bidirectionally (co-constructively) and across verbal and non-verbal channels. In our vision, the robot should anticipate explanation needs from the user, alleviating the cognitive burden of monitoring an unfamiliar explainee, by providing information about its understanding of the situation, of the next action required, or of the intended request by the user. Feedback and explaining cues are then provided in a multimodal way, also leveraging non-anthropomorphic features designed to be easily interpretable for novice users.

Study 1: Interactive task learning in Augmented Reality. In our first exploration, we targeted imitation learning in a virtual scene.  Naive users demonstrated kitchen tasks to a personal robot continuously learning from interaction. The robot signaled known objects and actions by displaying virtual labels on the shared workspace as the user demonstrated the task. It asked further questions to generalize the observed demonstration, also providing insight into its understanding. The user could further monitor the robot’s learning by asking it to plan a task based on the received training. This provided an easy scaffolding strategy for further demonstrations. Users found the system engaging, understandable, and trustworthy. Those who varied their demonstrations more were able to expand the robot’s knowledge more effectively, and those who felt they understood the robot better were also more trusting of it. 

Study 2: Attentive support in group interactions. In this study,  we leveraged the flexibility of Large Language Models (LLMs) 1) to understand natural language and reason about action needs in group scenarios, 2) to explain the robot’s reasoning and decision-making. The robot listens to the conversation between humans and carries out supportive actions if directly addressed or intervenes instead of another user if it detects that the addressee is currently hampered. Through a tailored prompt, the robot generates a verbal explanation for its behavior, demonstrating LLMs’ capability to integrate scene perception, dialogue acquisition, situation understanding, and behavior generation. 

Study 3: Mirror Eyes for intention recognition. While LLMs provide almost human-level natural language processing and generation, verbal explanations can be cumbersome and ambiguous - particularly in tasks involving spatial references and physical manipulation. In such cases, we often rely on non-verbal cues to monitor others’ understanding, primarily gaze direction. Head direction and 2D eye models often are inadequate for estimating the focus of attention of a robot in a cluttered scene. In this study , an interaction concept we refer to as Mirror Eyes  facilitated referential gaze understanding through the addition of a reflection-like image of the attended object on top of screen-based eyes. A user study demonstrated the benefit of this feature in a pick-and-place task, where users were able to detect robot errors earlier in comparison to nonreflective eyes and experienced a head with Mirror Eyes as the more explainable medium. 

Study 4: Gaze-based speech disambiguation. With a specular rationale to Study 3, we reasoned that for intuitive human-robot interaction, robots should be able to ground conversations by relating ambiguous or underspecified utterances to the current physical situation and to the intents expressed nonverbally by the user’s gaze. Our approach  relies on a text-based semantic translation of the scanpath produced by the user, along with the verbal request, and demonstrates LLM’s capabilities to reason about gaze behavior. Such representation allows resolving demonstrative disambiguation (e.g., ’give me that’) but also to accurately infer users’ intents by leveraging gaze to ground ambiguous speech and by focusing on speech content to discard spurious fixations.