PhysLDM: latent diffusion one-shot simulator predicts high-fidelity 3D deformable dynamics zero-shot
XingangP · x · 2026-10-19
Researchers at NTU S-Lab and Shanghai AI Lab introduce PhysLDM, a neural physics simulator built on latent diffusion that works 'like video generation' for physics.
- A spatiotemporal VAE compresses dynamic 3D deformable meshes into a compact latent space, avoiding the 'staircase' artifacts of video-style temporal compression and disentangling rigid motion from local deformation (2.48mm reconstruction accuracy at up to 78x token compression).
- Key insight: deformable dynamics are often chaotic, where deterministic regression yields damped, non-physical averages — diffusion better models the distribution, so the model predicts entire high-resolution trajectories in one shot.
- Trained purely kinematically on Objaverse-scale simulation data, a single model generalizes zero-shot to unseen GSO and Toys4K datasets.
- Fully differentiable, enabling challenging inverse problems.
The team frames it as a step toward physical world models.
More from Embodied
- Robots ace blind precision assembly using only joint encoder signals, no cameras — kevin_zakka · 2026-10-19
- If AI records every conversation by default, society changes dramatically — generativist · 2026-10-19
- Tesla shareholder letter teases a mysterious Optimus v3 silhouette — CyberRobooo · 2026-10-19
- Robotics is having its GPT-2 moment, says State of AI's Nathan Benaich — nathanbenaich · 2026-10-19
- Robotics founder's China day 6: not the top AI innovator, but adoption will lead by a massive gap — DominiqueCAPaul · 2026-10-19
- Ropedia Kicks Off Super Week With W0, a World Model Teaching Robots to Predict Physics Before Acting — liuziwei7 · 2026-10-19