Research

Memory Mechanisms in Spatial and Embodied Intelligence

Core Theme: Memory

In spatial intelligence, we study spatiotemporal reconstruction and persistent spatial memory. In embodied intelligence, we study pretrained foundation models, continual learning, reinforcement learning, and safe decision-making in agents and robotic systems.

Spatial Intelligence

Spatiotemporal reconstruction · Persistent spatial memory · Place recognition · Pose estimation

Spatial intelligence centers on persistent memory for understanding the world. We study how agents build and update spatial memory in open, dynamic environments over long periods, then use it for place recognition, long-term localization, large-scale scene understanding, and pose estimation.

Pose-Free Robust Spatiotemporal Generative Reconstruction

We study spatiotemporal reconstruction, sparse generative mapping, and pose estimation in dynamic, cluttered environments. The goal is to help agents build reliable spatial memories and reason about changing spatial states.

Pose-Free Streaming Spatial Reconstruction

We study streaming methods for spatiotemporal reconstruction, pose estimation, and multi-sensor fusion across large-scale scenes. These methods allow agents to build, maintain, and update spatial memory as new observations arrive.

Place Recognition and Persistent Spatial Memory

We study biologically inspired place memory and visual, LiDAR-based, and cross-modal place recognition. Our focus is on building, updating, and retrieving spatial memories, and on reliable relocalization as environments change over time.

Embodied Intelligence

Memory · Policy Models · Reinforcement Learning · Safe Decision-Making · Agents

Embodied intelligence focuses on embodied behavior and safe decision-making. We study how agents use evolving spatial memory to coordinate locomotion, perception, and manipulation over long-horizon tasks, while maintaining safety as they act.

Structured Memory Evolution for Agents

We study how agents can turn local observations into structured spatial memories that can be queried, updated, and reasoned over, supporting an understanding of the world throughout long-term tasks.

World Models, Vision-Language-Action Models, and Reinforcement Learning

We study how world models, vision-language-action models, and reinforcement learning can work together to help agents anticipate what may happen, interpret tasks, choose actions, and adapt to changing environments.

Safe Decision-Making

We study how historical experience, spatial context, and risk memory can support safe policy learning, anomaly avoidance, recovery from failure, and decisions that account for uncertainty.