During research voyages around Hawaii in 2018, Yuening Zhang SM ’19, PhD ’24, saw firsthand how challenging it was to keep a tight grip on a ship. The careful coordination required to map the underwater terrain could sometimes lead to a stressful environment for team members, who might have different understandings of what tasks needed to be completed in suddenly changing conditions. During these voyages, Zhang wondered how a robot companion could help him and his crew members achieve their goals more efficiently.
Six years later, as a research assistant at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), Zhang developed what could be called the missing piece: an AI assistant that could communicate with team members and coordinate roles to achieve a common goal. In a paper presented at the International Conference on Robotics and Automation (ICRA) and published August 8 in IEEE Xplore, Zhang and his colleagues present a system that can oversee teams of both humans and AI agents, intervening as needed to increase the effectiveness of teamwork in domains such as search and rescue missions, medical procedures, and strategy video games.
The CSAIL-led group developed a theory-of-mind model for AI agents that represents how humans think about and understand each other’s possible plans of action when collaborating on a task. By observing the behavior of its fellow agents, this new team coordinator can infer their plans and each other’s understanding from their prior sets of beliefs. If their plans are in conflict, the AI helper intervenes by reconciling each other’s beliefs, directing actions, and asking questions as needed.
For example, when rescuers triage victims on the scene, they have to make decisions based on their beliefs about each other’s roles and progress. This kind of epistemic planning can be improved by CSAIL’s software, which can send messages about what each agent is going to do or has done to ensure tasks are completed and avoid duplicated work. In this case, an AI helper can step in to tell agents that they have already progressed to a certain room, or that none of the agents have covered a particular area with a potential victim.
“Our work takes into account the sentiment: ‘I believe you believe the same thing as other people,'” says Zhang, now a research scientist at Mobi Systems. “Imagine working in a team. You ask yourself, ‘What on earth is that person doing? What am I going to do? Does he understand what I’m about to do?’ We model how different team members can understand the overall plan and communicate what they need to accomplish to achieve the team’s overall goal.”
AI is the savior
Even with careful planning, human and robotic agents can get confused and make mistakes if their roles aren’t clear. This predicament is especially acute in search and rescue missions, where the goal is to find someone in danger while scanning a vast area in a limited time. Thankfully, enhanced communications technology in the new robot assistants could inform search teams of what each group is doing and what they’re searching for, allowing agents to navigate the terrain more efficiently.
This type of task organization could also be useful in other high-risk scenarios like surgery, where a nurse first takes the patient to the operating room, then an anesthesiologist puts the patient to sleep before the surgeon starts the operation. During the operation, the team needs to continuously monitor the patient’s condition while dynamically responding to each colleague’s actions. To make sure each activity during the operation is properly organized, an AI team coordinator can monitor and intervene if any of these tasks become disruptive.
Effective teamwork is also essential in video games like “Valorant,” where players work together to coordinate who should attack and who should defend against other teams online. In such scenarios, an AI assistant can pop up on the screen and warn individuals where they’ve misunderstood the task they need to complete.
Prior to leading the development of this model, Zhang designed a computational model, EPike, that could act as a team member. In a 3D simulation program, the algorithm controlled a robotic agent that had to match a container to a drink selected by a human. No matter how rational and sophisticated these AI-simulated bots are, they can be limited by misunderstandings about their human partners and the task. The new AI coordinator can correct the agent’s thinking as needed to solve potential problems, and it did so consistently in this example. The system sent messages to the robot about the human’s true intentions so that the robot could match the container correctly.
“Over the years, in our work on human-robot collaboration, we’ve been humbled and inspired by how fluid human partners can be,” says Brian C. Williams, MIT professor of aeronautics and astronautics, CSAIL member, and senior author of the study. “Think of a young couple with their children. They work together to make breakfast for the kids and get them off to school. If one parent sees their partner making breakfast and still in their bathrobe, they’ll know, without a word, to rush off to shower and get the kids off to school. Good partners are well attuned to each other’s beliefs and goals, and our work on epistemic planning aims to capture this type of reasoning.”
The researchers’ method integrates probabilistic reasoning and recursive mental modeling of agents, enabling AI assistants to make risk-limited decisions. Moreover, they focus on modeling the agent’s understanding of plans and actions, which could complement previous work on modeling beliefs about the current world and environment. While AI assistants currently infer an agent’s beliefs based on prior information of possible beliefs, the MIT group envisions applying machine learning techniques to generate new hypotheses on the fly. To apply this counterpart to real-world tasks, they aim to consider richer plan representations in their work and further reduce the computational cost.
Dynamic Object Language Labs president Paul Robertson, Johns Hopkins assistant professor Tianmin Shu, and formerly of CSAIL Sungkweon Hong PhD ’23 joined Zhang and Williams on the paper. Their work was supported in part by the Defense Advanced Research Projects Agency’s (DARPA) Artificial Social Intelligence for Successful Teams (ASIST) program.
