Rapid trial-and-error learning with simulation supports flexible tool use and physical reasoning

TL;DR

This work investigates rapid, flexible physical problem solving and tool use by comparing human performance to a minimal, internal model-driven agent. It presents the Virtual Tools game and the SSUP framework, consisting of Sample (object-based priors), Simulate (noisy forward physics), and Update (learning from thoughts and actions) to explain how people plan and adapt in novel tasks. Across 30 levels and a validation set, the SSUP model matches human solution rates and action sequences, with ablations showing each component's importance; the novel-level data show strong correlations with human performance (e.g., r=0.85 for solution rates and r=0.95 for accuracy). The framework offers a general approach to condense rich world knowledge into task-specific plans, highlighting avenues for more sophisticated priors and forward-planning in AI that reasons about physical dynamics.

Abstract

Many animals, and an increasing number of artificial agents, display sophisticated capabilities to perceive and manipulate objects. But human beings remain distinctive in their capacity for flexible, creative tool use -- using objects in new ways to act on the world, achieve a goal, or solve a problem. To study this type of general physical problem solving, we introduce the Virtual Tools game. In this game, people solve a large range of challenging physical puzzles in just a handful of attempts. We propose that the flexibility of human physical problem solving rests on an ability to imagine the effects of hypothesized actions, while the efficiency of human search arises from rich action priors which are updated via observations of the world. We instantiate these components in the "Sample, Simulate, Update" (SSUP) model and show that it captures human performance across 30 levels of the Virtual Tools game. More broadly, this model provides a mechanism for explaining how people condense general physical knowledge into actionable, task-specific plans to achieve flexible and efficient physical problem-solving.

Figures (8)

• Figure 1: Examples of using objects to achieve a goal. (A) Bearded capuchin monkey opening a cashew nut with an appropriately sized stone luncz2016wild. (B) New Caledonian crow using heavy blocks to raise the water level in a tube in order to retrieve food jelbert2014using. (C) Toddler using a shovel to reach a ball (from https://youtu.be/hwrNQ93-568?t=198). (D) One illustrative trial in the Virtual Tools game (https://sites.google.com/view/virtualtoolsgame). (i) The player must get the red object into the green goal using one of the three tools. (ii) The player chooses a tool and where to place it. (iii) Physics is turned "on" and the tool interacts with other objects. The action results in a near miss.
• Figure 2: Twenty levels used in the Virtual Tools game. Players choose one of three tools (shown to the right of each level) to place in the scene in order to get a red object into the green goal area. Black objects are fixed, while blue objects also move; grey regions are prohibited for tool placement. Levels denoted with A/B labels are matched pairs.
• Figure 3: Examples of participants' behavior on three levels, representative of rapid trial-and-error learning: Initial plans are structured around objects, followed by exploring to identify more promising strategies and then refining actions until success. Objects start as shown by light blue/red outlines and follow paths traced out by colored lines. Possible tool choices shown to the right. (A) In the Catapult level, a useful strategy is often identified immediately and rapidly fine-tuned. (B) Other participants first try an unsuccessful strategy but then switch to a more viable strategy and refine it. (C) The Launch (B) level is designed to prevent obvious solutions. This participant may have initially believed the ball would start rolling and attempted to use a tool as a bridge. When this failed they realized they needed to launch the ball, but only discovered after several trials how to use a tool in a non-obvious way to accomplish this, via a hooking motion around the blocking ledge. They then took several more trials to fine-tune this action. (D) In the SeeSaw level, a participant realized on the second attempt they must support the platform for the ball to roll across, then tried different ways of making this happen.
• Figure 4: (A) The SSUP algorithm. (B) A diagram of the model for the Virtual Tools game. It incorporates an object-based prior, a simulation engine for filtering proposals, and an update module that suggests new proposals based on observations "in the mind" and from actions taken in the world. (C) Illustration of the policy $\pi'$ evolving while attempting a level. Colored patches represent the Gaussian policy for each tool.
• Figure 5: (A) Comparison of average number of human participants' attempts for each level with average number of attempts for the SSUP model. Bars indicate $95\%$ confidence intervals on estimates of the means. (B) Comparison of human participants' accuracy on each trial versus the accuracy of the SSUP model. (C) Comparison of human participants' accuracy to all alternate models. Numbers correspond to the trials in Fig. \ref{['fig:trials']}.