Research map / Multi-agent systems and games / Human-AI and goal-conditioned
I-PHYRE: Interactive Physical Reasoning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces I-PHYRE, a 2D block-elimination benchmark of 40 games in four splits that tests interactive physical reasoning (intuitive physics, multi-step planning, in-situ timing), and finds RL and supervised agents lag far behind humans in zero-shot generalization.
Problem
Existing physical reasoning benchmarks evaluate stationary scenes or single interventions, so they cannot assess multi-step interactive physical reasoning with precise action timing in dynamic scenes.
Approach
A pymunk/pygame block-elimination environment where agents remove blocks at chosen times, evaluated under three planning strategies (planning in advance, planning on-the-fly, combined) and trained on a basic split then zero-shot tested on noisy, compositional, and multi-ball splits.
Key results
- Human participants achieved an 80%+ success rate across splits (basic 92.39%, noisy 90.00%, compositional 82.83%, multi-ball 82.83%) versus oracle 100%.
- Supervised classifiers (Global, Object, Vision Fusion) reached roughly 54-60% accuracy on generalization splits, close to random guess (~50%).
- Best RL agent PPO-I averaged 528.10 reward across 40 games with 57.5% success rate, while human average reward was 844.17 with 87.55% success rate.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use planning in advance instead of planning on-the-fly | PPO on I-PHYRE · I-PHYRE 40-game average | average reward: 377.74 -> 528.10 | Helped |
| use the combined strategy instead of planning on-the-fly | PPO on I-PHYRE · I-PHYRE 40-game average | average reward: 377.74 -> 504.33 | Helped |
| use planning in advance instead of the combined strategy | PPO on I-PHYRE · I-PHYRE 40-game average | average reward: 504.33 -> 528.10 | Helped |
| use planning in advance instead of planning on-the-fly | A2C on I-PHYRE · I-PHYRE 40-game average | average reward: 352.69 -> 429.36 | Helped |
| add pre-trained ViT visual features to Global Fusion | Supervised learner on I-PHYRE · I-PHYRE noisy | accuracy: 59.8 -> 57.5 | Hurt |
| add pre-trained ViT visual features to Global Fusion | Supervised learner on I-PHYRE · I-PHYRE compositional | accuracy: 57.0 -> 55.8 | Hurt |
| add pre-trained ViT visual features to Global Fusion | Supervised learner on I-PHYRE · I-PHYRE multi-ball | accuracy: 56.7 -> 59.5 | Helped |
| use Object Fusion rather than Global Fusion | Supervised learner on I-PHYRE · I-PHYRE compositional | accuracy: 57.0 -> 60.2 | Helped |
| train on the basic split and evaluate on similar unseen basic games | A2C-I · I-PHYRE unseen basic | average reward: 862.47 -> 862.25 | No clear effect |
| use a model-based World Model alongside PPO | PPO on I-PHYRE · I-PHYRE generalization splits | performance | No clear effect |
| use the supervised classifier trained on naively sampled action sequences | Supervised learners on I-PHYRE · I-PHYRE generalization splits | accuracy | No clear effect |
Surprising findings
- Supervised learners trained on the basic split barely outperform random guessing on generalization splits.
- GPT-4 with prompting obtains a -29.10 average reward on the compositional split despite succeeding on basic and multi-ball splits.
Limitations
- Environment restricted to 2D physics with 40 games.
- RL agents struggle with sparse no-op-dominated action distributions and precise timing.
- Offline methods (Decision Transformer, World Model) fail to improve over model-free RL.
Experimental setup
Datasets: I-PHYRE basic split I-PHYRE noisy split I-PHYRE compositional split I-PHYRE multi-ball split
Benchmarks: I-PHYRE
Metrics: average reward success rate mean accuracy (supervised) iterations to sample 50 successful sequences
Baselines: PPO A2C SAC DDPG DQN World Model Decision Transformer Global Fusion Object Fusion Vision Fusion GPT-4 Human participants
Builds on: PHYRE World Models Decision Transformer PPO
Closest papers by meaning
- Causal-PIK: Causality-based Physical Reasoning with a Physics-Informed Kernel (ICML 2025 · similarity 0.87)
- PhyBlock: A Progressive Benchmark for Physical Understanding and Planning via 3D Block Assembly (NeurIPS 2025 · similarity 0.86)
- PhysVLM-AVR: Active Visual Reasoning for Multimodal Large Language Models in Physical Environments (NeurIPS 2025 · similarity 0.86)
- PhysGym: Benchmarking LLMs in Interactive Physics Discovery with Controlled Priors (NeurIPS 2025 · similarity 0.86)
- PhysBench: Benchmarking and Enhancing Vision-Language Models for Physical World Understanding (ICLR 2025 · similarity 0.86)
- Kinetix: Investigating the Training of General Agents through Open-Ended Physics-Based Control Tasks (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.