We believe that grounding can be seen as a chain-of-thought for spatial reasoning. Based on this, we achieve a new SOTA performance on VSI-Bench.

We present MM-Nav a multi-view VLA system with 360° perception. The model is trained on large-scale expert navigation data collected from multiple reinforcement learning agents, demonstrating robust generalization in complex navigation scenarios.

We introduce the concept of semantic orientation, representing the object orientation condition on open vocabulary language.

We recast the vision–language–action model as a perception–prediction–action model and make the model explicitly predict a compact set of dynamic, spatial and high- level semantic information, supplying concise yet comprehensive look-ahead cues for planning.

We introduce DexVLG, a vision‑language‑grasp model trained on the 170M‑pose, 174k‑object dataset that can generate instruction‑aligned dexterous grasp poses and achieves SOTA success and part‑grasp accuracy.

We introduce Hybrid‑depth, a self‑supervised method that aligns hybrid semantics via language guided fusion, achieving SOTA accuracy on KITTI and boosting downstream perception.

We rethink the role of positional encoding in 3D representation learning, and propose Positional Prompt Tuning, a simple but efficient method for transfer learning.

Based on cognitive psychology, we introduce a comprehensive and complex spatial reasoning benchmark, including 50 detailed categories and 1.5K manual labeled QA pairs.

We collect diverse images and prompts, and utilize GPT-4o for automated evaluation aligned with human preference.

We present ShapeLLM, the first 3D Multimodal Large Language Model designed for embodied interaction, exploring a universal 3D object understanding with 3D point clouds and languages.

We enhance the utilization of color information to improve 3D scene self-supervised learning.

We present DreamLLM, a learning framework that first achieves versatile Multimodal Large Language Models empowered with frequently overlooked synergy between multimodal comprehension and creation.

We achieve rapid, multi-category 3D conditional generation by sharing the merits of different representations. VPP can generate 3D shapes less than 0.2s using a single RTX 2080Ti.

We propose contrast guided by reconstruct to mitigate the pattern differences between two self-supervised paradigms.

We propose to use autoencoders as cross-modal teachers to transfer dark knowledge into 3D representation learning.