ADELPHI, Md. -- Army researchers are working to improve Soldier-robot teaming in tactical environments by enabling robots to ask questions and learn in real-time through dialogue.
According to Dr. Felix Gervits, researcher at the U.S. Army Combat Capabilities Development Command, known as DEVCOM, Army Research Laboratory, tactical environments for Soldier-robot teaming are characterized by high degrees of novelty and uncertainty.
Robots operating in such environments will naturally encounter unfamiliar concepts such as objects and locations and will need methods to learn what these are, he said.
One solution to this problem is to leverage the Soldier, and allow robots to ask questions to support learning; however, it is not clear what kinds of questions robots should ask, nor how they should phrase the questions, Gervits said.
Researchers transcribed and labeled the dialogues from the study for various question types, and then publicly released as the Human-Robot Dialogue Learning, or HuRDL, corpus.
This research has three main contributions:
- Software platform for online experimentation
- Labeling scheme for categorizing questions
- Labeled HuRDL corpus
As a step toward enhancing Soldier-robot dialogue, DEVCOM ARL researchers, in collaboration with the Naval Research Laboratory and Tufts University, collected dialogue data from a human-robot interaction study to investigate how robots should ask questions when confronted with novel concepts.
This study is the first part of a broader research effort to develop algorithms to support automated question generation and learning for Soldier-robot teaming, Gervits said.
“The broad goal of this research is to improve Soldier-robot teaming in tactical environments by enabling robots to ask questions and learn in real-time through dialogue,” Gervits said. “The current research is a step toward this goal because it highlights the kinds of questions that people ask when encountering unfamiliar concepts. The data that we collected in the HuRDL corpus can be used to train robots to ask specific questions when they encounter similar kinds of uncertainty. They can then learn from the responses or ask follow up questions if necessary.”
One potential issue with learning through dialogue is that the robot can ask many questions and cause frustration or disruption to the team.
This research seeks to avoid this problem by enabling the robot to ask the fewest, most efficient questions needed to learn a concept, Gervits said. By analyzing effective dialogue strategies and question types used in the study, they can implement policies to ensure that robots cause minimal disruption while learning.
To conduct interactive studies in the midst of the COVID-19 pandemic, Gervits and colleagues developed a software platform for online experimentation. The platform allows for crowdsourced data collection in which people from all over the country can participate remotely in virtual ARL experiments.
The current study involved participants controlling a robot in a virtual 3D environment that contained a variety of unfamiliar objects and locations.
According to Gervits, back-and-forth communication was at the heart of the study. An experimenter gave instructions to the participants to locate and move some of the objects. Because the objects were unfamiliar and had strange names and properties, people tended to ask many questions. These questions were recorded for later analysis.
Analysis of the questions from the study led to the creation of what is known as an annotation scheme, or a method to structure dialogue data.
The scheme categorized questions based on their form and function.
For example, researchers labeled the question “Is it the red one?” as a yes-no question of the type “confirm object by color.”
“This way of rigorously structuring dialogue data is an essential step in the process of developing automated approaches to question generation for robot learning through dialogue,” Gervits said.
This annotation scheme was applied to the HuRDL corpus and every question was labeled accordingly.
According to the researchers, this research stands out in several ways from the field.
First, it emphasizes back-and-forth dialogue as a means of learning, which is different from traditional active learning approaches, which generally focus on one-off queries or non-linguistic methods such as learning from demonstration. Since language is a natural interface for humans, this approach is intuitive and requires minimal training.
Second, the research is grounded in empirical, interactive studies of human dialogue. In contrast with typical online studies in which participants view pictures or videos and answer survey questions, this study involved a real-time interaction between experimenter and participant. This was enabled by the team’s novel software framework for online experimentation.
Finally, the study environment the researchers used was an Army-relevant one in which people were situated in an unexplored area and relied on a variety of perceptual properties (object features, landmarks, etc.) to learn the names of novel objects. As a result, the findings of the study should apply broadly to Army-relevant environments.
Moving forward, the labeled HuRDL corpus will serve as a useful resource for researchers studying human-robot dialogue in situated interaction.
Additional analysis of the corpus will reveal dialogue strategies for effective question generation and other relevant findings such as the best time to ask questions and how many questions are appropriate, Gervits said.
Gervits noted that the main direction for future research is the creation of novel algorithms for automated question generation based on the data collected in the study.
The data can either be used directly to train an automated system, or to inform general policies for question generation. Either approach will get the team closer to the goal of having robots that can ask effective questions and learn from the answers, he said.
“As teams of Soldiers and robots become increasingly common in the military, it will be necessary to coordinate the activity of these groups,” Gervits said. “Dialogue is a natural and effective way of achieving this goal, and the current research is a step in the direction of creating a natural speech interface that allows robots to learn about the world through dialogue with Soldiers. I am optimistic that this research will have a real impact on the future of Soldier-robot teams in the Army.”
This research will be presented at the Special Interest Group on Discourse and Dialogue 2021 conference by the team including Gervits, ARL’s Dr. Matthew Marge, Dr. Gordon Briggs at NRL, and Drs. Antonio Roque and Matthias Scheutz at the Tufts University Human-Robot Interaction Lab. The research will also be published in the conference proceedings.
The HuRDL corpus is available online.
As the Army’s national research laboratory, ARL is operationalizing science to achieve transformational overmatch. Through collaboration across the command’s core technical competencies, DEVCOM leads in the discovery, development and delivery of the technology-based capabilities required to make Soldiers more successful at winning the nation’s wars and come home safely. DEVCOM Army Research Laboratory is an element of the U.S. Army Combat Capabilities Development Command. DEVCOM is a major subordinate command of the Army Futures Command.