Causal-PIK: Causality-based Physical Reasoning with a Physics-Informed Kernel

Tasks that involve complex interactions between objects with unknown dynamics make planning before execution difficult. These tasks require agents to iteratively improve their actions after actively exploring causes and effects in the environment. For these type of tasks, we propose Causal-PIK, a method that leverages Bayesian optimization to reason about causal interactions via a Physics-Informed Kernel to help guide efficient search for the best next action. Experimental results on Virtual Tools and PHYRE physical reasoning benchmarks show that Causal-PIK outperforms state-of-the-art results, requiring fewer actions to reach the goal. We also compare Causal-PIK to human studies, including results from a new user study we conducted on the PHYRE benchmark. We find that Causal-PIK remains competitive on tasks that are very challenging, even for human problem-solvers.

Humans are skilled at solving tasks that involve complex object interactions—such as dropping an object onto one end of a lever to create a catapult effect, launching another object from the opposite end. These tasks pose a major challenge for artificial intelligence systems, as the outcomes of actions depend on complex and often unknown physical dynamics. Humans approach these problems with physical intuition, learning quickly from previous attempts to refine their strategies.Inspired by this ability, we developed Causal-PIK, a method that helps artificial agents to solve physical reasoning tasks more efficiently by reasoning about object interactions. Just like how humans efficiently learn from their mistakes when solving physical problems, our method allows agents to learn from their mistakes and make increasingly informed decisions with fewer trials. The key innovation lies in the algorithm's ability to understand direct cause-and-effect relationships and perform counterfactual reasoning — allowing it to imagine alternate scenarios and evaluate what could have happened under different conditions. Our findings show that the ability to predict which actions will cause similar effects significantly impacts an agent’s efficiency in exploring complex environments.