Search papers, labs, and topics across Lattice.
7
0
7
Achieve real-time 360掳 robotic vision with RobotPan, which drastically reduces the number of Gaussians needed for reconstruction and view synthesis compared to previous feed-forward methods.
Forget finetuning: this training-free method achieves state-of-the-art zero-shot 3D visual grounding, even in messy, real-world environments.
Sparse-view 3D Gaussian Splatting gets a major boost by incorporating priors from geometry foundation models and VLMs to overcome the limitations of color residual heuristics.
Ditch VAEs and AMP: SLMP learns structured motion priors in a spherical latent space, enabling stable random sampling of diverse and valid humanoid behaviors without information loss.
Turns out, robotic manipulation policies struggle with tasks requiring memory, and this benchmark reveals the architectural design choices that actually matter for improving performance.
Generate superhuman humanoid motion data at scale by closing the loop between policy performance and data difficulty, surpassing limitations of fixed datasets.
Humanoid robots can now learn complex, terrain-aware motions directly from video using a low-cost pipeline, eliminating the need for expensive MoCap data and manual motion design.
