I-PHYRE: Interactive Physical Reasoning

TL;DR

This work introduces I-PHYRE, a framework that challenges agents to simultaneously exhibit intuitive physical reasoning, multi-step planning, and in-situ intervention, and examines several supervised and reinforcement agents' zero-shot generalization proficiency on this framework.

Abstract

Current evaluation protocols predominantly assess physical reasoning in stationary scenes, creating a gap in evaluating agents' abilities to interact with dynamic events. While contemporary methods allow agents to modify initial scene configurations and observe consequences, they lack the capability to interact with events in real time. To address this, we introduce I-PHYRE, a framework that challenges agents to simultaneously exhibit intuitive physical reasoning, multi-step planning, and in-situ intervention. Here, intuitive physical reasoning refers to a quick, approximate understanding of physics to address complex problems; multi-step denotes the need for extensive sequence planning in I-PHYRE, considering each intervention can significantly alter subsequent choices; and in-situ implies the necessity for timely object manipulation within a scene, where minor timing deviations can result in task failure. We formulate four game splits to scrutinize agents' learning and generalization of essential principles of interactive physical reasoning, fostering learning through interaction with representative scenarios. Our exploration involves three planning strategies and examines several supervised and reinforcement agents' zero-shot generalization proficiency on I-PHYRE. The outcomes highlight a notable gap between existing learning algorithms and human performance, emphasizing the imperative for more research in enhancing agents with interactive physical reasoning capabilities. The environment and baselines will be made publicly available.

Figures (10)

• Figure 1: Illustrations of the four splits in benchmark. The objective is to guide all red balls into the hole by strategically eliminating gray blocks in a stepwise manner. We showcase one game from each split (basic, noisy, compositional, and multi-ball) and explore the potential dynamics that different interventions could unfold. Images encased in a black border, cyan border, green border, and red border represent intermediate states, time steps at which an elimination action is undertaken, frames depicting success, and frames indicating failure, respectively. The gray and black blocks remain stationary, while the light blue blocks are mobile, influenced by gravity and collisions; only the gray blocks are subject to elimination. Successfully solving the games necessitates reasoning about the sequence of interventions and meticulously managing the exact timing of interventions.
• Figure 2: Performance of various rl agents on benchmark. Agents, trained on the basic split, are evaluated on the remaining three splits in a zero-shot fashion. The suffixes '-I', '-O', and '-C' denote planning in advance, on-the-fly planning, and the combined strategy, respectively. The dashed lines are human results.
• Figure 3: The training curves of different rl agents on the basic split. Training steps vary among agents.
• Figure A1: The iteration numbers required to generate 50 different successful action sequences.
• Figure A2: The initial scenes of basic games.