A research team has built a physics dynamics model that handles collisions, deformation, and energy loss together — instead of bolting on each effect separately.
The system, called NEXUS, represents objects as structural graphs and tracks how they interact through contact. Rather than predicting positions or accelerations directly, it models motion using scalar energy and dissipation terms borrowed from Hamiltonian mechanics. Gravity and elastic deformation stack as additive energy terms; damping and impact losses get their own learned dissipation functions. Forces are then derived by differentiating those functions, and the whole thing is integrated forward with a multi-substep semi-implicit solver — a standard trick for keeping simulations stable at large time steps.
The reason this matters is that video generation models currently fake physics rather than simulate it. They learn statistical correlations between frames, which means a bouncing ball or a crumpling can looks plausible for a second or two before going subtly wrong. NEXUS trajectories serve as scaffolding: feed them into a video generator and the output stays physically consistent over longer horizons without sacrificing visual quality.
On trajectory benchmarks, NEXUS outperforms both pure learned baselines and hybrid physics-structured approaches across varied material properties and force combinations — though benchmarks designed by the same team that built the system are always worth reading with a degree of skepticism. The harder test will come when someone runs it against scenes the authors did not tune for.