This work presents a few-shot learning framework for planar pushing of novel objects with unknown physical properties, including center of mass (CoM), friction, and inertia.

We introduce SimPush, a large-scale simulation dataset containing over two million pushes across diverse object shapes and physical configurations.

Based on this dataset, we develop an attention-based encoder–decoder model that leverages a small number of prior push interactions to predict object motion (Δx, Δy, Δθ) together with uncertainty estimates.

Building on the predicted push affordances, we propose an uncertainty-aware planning strategy that efficiently selects pushing actions to translate objects to target poses.

The method achieves accurate motion prediction in simulation and successfully transfers to real-world experiments without real-data training, while requiring fewer push steps than baseline methods.