Search papers, labs, and topics across Lattice.
5
0
5
GARFIELD enables real-time, uncertainty-aware motion planning by efficiently modeling the distribution of possible scene futures, outperforming traditional methods by a factor of 97 in trajectory sampling speed.
Contrastive supervision can unlock implicit disentanglement in generative models, leading to superior content-style separation and robustness against distribution shifts.
Forget generating entire videos – this method distills motion into a highly compressed latent space, letting you steer scene dynamics with text prompts at unprecedented speeds.
Imagine simulating thousands of plausible futures from a single image, guided by motion constraints, with accuracy rivaling dense simulators but at warp speed.
By encoding objects as local 3D meshes, FlowTouch achieves view-invariant visuo-tactile prediction that generalizes across sim-to-real and different sensor instances, a leap beyond camera-dependent mappings.
