Generating Pattern-Based Conventions for Predictable Planning in Human-Robot Collaboration

Clare Lohrmann, Maria Stull, Alessandro Roncone, Bradley Hayes

When humans interact with robots, they often find them to be unpredictable - not matching their expectations. This results in a lack of trust, understanding, as well as negative perceptions of the robot and difficulty collaborating. Often when the robot's behavior is unpredictable, attempts are made to explain the robot's behavior to humans. This requires the robot to provide explanations that identify the mismatch in human expectation and ground truth, as well as construct an explanation that fully rectifies this mismatch - a tall order. Our work takes a different approach and addresses the robot's behavior directly. We improve the predictability of the robot by leaning into human cognitive tendencies and producing robot behavior that follows a human-legible pattern-based convention. Our work improves robot predictability by constraining robot behavior to be in line with pattern recognition and abstraction processes that human brains are built for.

We present an algorithmic approach to select a pattern-based behavioral convention for the robot to adhere to. We utilize a predictability metric that constructs patterns from human-legible features of the subtask space, and scores them on determinism and uniqueness to select the ideal pattern convention for the environment. The use of this metric, which we call PACT, results in better coordination within a human-robot dyad in a collaborative game environment, as well as improves human perceptions of robot predictability, understandability, and the team itself.

The full paper can be accessed here and from our Publications tab. A blog-post with further details can be accessed here