📌 Catch our poster at ICML 2025
🗓️ Thursday, July 17th, 11:00 AM–1:30 PM PDT
📍 East Exhibition Hall A-B, Booth E-2107

For more details on our work, please visit our project website 🔗: stanford-iprl-lab.github.io/causal-pik/

🎉 Huge thanks to the amazing team Michelle Yi, Claire Chen, Sarah A. Wu, Rika Antonova, @tobigerstenberg.bsky.social, and @leto--jean.bsky.social

✅ Causal-PIK outperforms state-of-the-art baselines, requiring fewer attempts to solve the tasks.
✅ High correlation in scores between humans and Causal-PIK suggests high alignment in the types of physical dynamics that were found to be easy or difficult to reason about.

(6/6)

📊 We evaluate our method on Virtual Tools and PHYRE, benchmarks designed to evaluate agents on single-intervention tasks that involve complex physical interactions between objects.

(5/6)

2️⃣ Smart action selection - Causal-PIK separates the direct effects of an action from confounding factors arising from the environment, allowing informed decision-making.

(4/6)

1️⃣ Learning from fewer examples — Causal-PIK generalizes from a single experience to multiple actions by predicting which ones will lead to similar outcomes.

(3/6)

🧩 We introduce Causal-PIK, a new method that combines Bayesian Optimization with a Physics-Informed Kernel to solve tasks with complex dynamics.

At its core, it uses causal reasoning, enabling:

(2/6)