ODeform uses Neural ODEs for continuous 4D shape deformation

Modeling how objects deform over time is crucial for robotics and simulation, but existing methods either rely on discrete time steps (losing continuity) or are too slow for real-time use. ODeform, accepted at IROS 2026, solves this by applying Neural Ordinary Differential Equations to 4D deformation. It takes 3D point clouds and physical conditions (like material properties), maps them into a unified latent space, and solves the resulting ODE to predict continuous motion flows. This bypasses the need for frame-by-frame processing, enabling smooth deformation trajectories at lower computational cost.

The researchers tested ODeform on unseen physical parameter configurations, showing improved motion prediction accuracy over baseline methods. They also demonstrated successful transfer to real 3D captured objects with novel shapes, and effective interpolation/extrapolation of learned dynamics. The approach's efficiency and continuity make it suitable for real-time applications like robotic manipulation, where objects change shape unpredictably. Code and data will be released, opening up further research in continuous physics-based simulation.