Proceedings of the Annual Meeting of the Cognitive Science Society

To support the systematic study of long-horizon human strategies in physical problem solving, we introduce a set of 28 virtual physics puzzles that require sequential manipulation and involve diverse interaction mechanics such as pick-and-place, pushing, and tool use and creation. The paradigm provides rich interaction logs and object-state trajectories, enabling computational analyses of strategies. We analyze data from 38 participants and estimate puzzle difficulty and participant performance. Using trajectory-based clustering, we identify recurring solution strategies and quantify solution diversity and repeatability between attempts. The results show that puzzles requiring both tool creation and pushing are the most challenging to solve with the longest time-to-solution and the highest failure rates, as they combine the large search spaces of tool use with the challenges of precise control and less predictable dynamics. Notably, participants demonstrate strong improvement after solving the puzzles once, suggesting learning and reuse of discovered strategies.