Skip to content

Event

PhysLDM: Latent Diffusion for Physical Simulation

1 reports · 1 sources

MMLab at Nanyang Technological University and Shanghai AI Laboratory propose PhysLDM, which uses a spatiotemporal VAE to compress entire trajectories of deformable object simulations into a latent space, then generates them in one shot with a diffusion model, achieving an average reconstruction error of 2.48 millimeters. The team used the GPU finite element solver StiffGIPC to generate about 80,000 simulation trajectories on more than 30,000 Objaverse objects for training. It is claimed that after training on Objaverse, when applied directly to unseen GSO and Toys4K objects, the reconstruction error remains at the same level as within the training distribution; the model can be extended to multi-object scenes without changing its architecture, and it is end-to-end differentiable.

Report timeline

  1. NTU MMLab and Shanghai AI Lab propose PhysLDM: high-fidelity physics simulation with latent diffusion

    NTU MMLab and Shanghai AI Lab propose PhysLDM, which uses a spatiotemporal VAE to compress an entire deformable-object simulation trajectory into latent space and then generates it in one pass with a diffusion model, achieving an average reconstruction error of 2.48 millimeters.

    机器之心 · 微信公众号 ·

Related topics

Shanghai AI Laboratory - Embodied IntelligenceRobot models and learningData and simulationResearch papers

Interest trend

Current
3
Peak
6Nov 4 01:00
Past 24 hours
-50%

The trend compares participants observed throughout the same period, so its range may be smaller than the current interest count. Move or tap to inspect each hour; use the arrow keys on a keyboard.