Event
GeoHAT Mobile Manipulation Diffusion Policy Paper Released
1 reports · 1 sources
Xiangyu Zhu and other authors have released GeoHAT, an end-to-end diffusion policy framework for mobile manipulation. It uses a lightweight Fourier spatial encoder to map per-pixel 3D coordinates into geometric tokens, which are then injected into vision foundation model features gated by depth validity, without requiring an additional 3D vision backbone. The authors report that on the ManiSkill-HAB simulation benchmark, GeoHAT achieves an average success rate of 79.2%, 23.6 percentage points higher than the strongest baseline; on multiple real-world tasks, it also shows consistent improvements over all baselines. These results are as stated by the paper's authors.
Report timeline
- GeoHAT: Geometry-Adaptive Hybrid Action Transformer for Mobile Manipulation
GeoHAT is an end-to-end diffusion policy framework that uses a lightweight Fourier spatial encoder to map per-pixel 3D coordinates into geometric tokens, then injects them into vision foundation model features via depth-validity-gated fusion, without requiring an additional 3D vision backbone.
arXiv cs.RO ·