Enhancing Robotic Competitors through Dynamic Strategy Adaptation to Humans

In this work, we are interested in making robots better at being competitive adversaries to human players. Robots can achieve that by effectively comprehending and dynamically adapting their behavior to the ever-changing strategies employed by human players in a game. We demonstrate that in the context of the game of tag, where a human-controlled robot and AI-controlled robot play the game multiple times with each other.

To enable a robot to dynamically adapt to changes in the human’s strategies, we first need to estimate the current strategy of the human player. To achieve this, we construct a set of reasonable strategies by modeling human behavior within the known game rules or by using machine learning techniques to replicate human-like gameplay. Subsequently, we employ Bayesian filtering or supervised learning methods to refine our beliefs regarding the human player’s strategy iteratively. To reduce the AI agent’s uncertainty regarding the human player’s strategy, the AI-controlled robot could strategically influence the human player’s actions, prompting them to reveal their ongoing strategy. This is achieved by directing them to specific areas within the game environment where their actions reveal crucial insights into their strategy. We call those points in the environment critical decision points. Therefore, the AI agent must take actions that steer the human agent towards those critical decision points. However, while steering the human agent to the critical decision point, to be an effective competitor, the AI agent must also be mindful of satisfying its objectives in the game, which is adversarial to the human agent. Therefore, it needs to balance the two objectives of influencing the human agent to reveal their strategy while optimizing the game’s winning.