Research agenda
Active embodied intelligence is an evidence problem.
I study active embodied intelligence under partial observability, with a focus on active perception, world models, and self-improving agents. At Avant Robotics, I develop outdoor robot simulation infrastructure and cross-embodiment long-range navigation models.
研究不完整观测下的主动具身智能，重点关注主动感知、世界模型与自进化智能体；在 Avant Robotics，主要开发户外机器人仿真基础设施与跨本体长程导航模型。
Active Perception for UAV Inspection
A drone rarely begins with all the evidence an inspection task requires. The research question is how to coordinate wide-field sensing, a high-resolution gimbal, and body motion while respecting viewpoint, bandwidth, latency, and safety constraints.
The goal is a policy that can decide what evidence is missing, where to acquire it, and when the evidence is sufficient. This is a research agenda, not a claim of a completed benchmark or deployed system.
Outdoor Simulation and Data Infrastructure
At Avant Robotics, I develop Avant-AGSim, an outdoor navigation simulation engine built on UE5, CARLA, and ProjectAirSim. It integrates scenes, dynamics, sensors, and control interfaces for aerial and ground robots in a shared environment, with reusable scene and task configurations.
My work covers automated collection of synchronized observations, trajectories, and scene annotations, target-visibility checks, and data-quality filtering. Validation includes 800 synchronized aerial–ground sample groups, rejection of seven categories of data anomalies, 1,000 repeated loading tests, and extended stability testing.
Independent robot control, synchronized observations, fault isolation, and recording and replay support multi-robot task development and repeatable data collection.
Cross-embodiment Long-range Navigation
I am developing a 4B vision-language navigation model using language instructions, first-person observation histories, and trajectory actions, with simulation data spanning UAVs, quadrupeds, wheeled robots, and humanoids.
This ongoing research studies how to reuse semantic and place memory across embodiments while accounting for differences in body size, motion capabilities, and obstacles. I design methods for observation-history compression, task-progress tracking, and failure recovery through blockage detection, alternative-route planning, and memory updates.
Outdoor tasks range from 350 m to 1.5 km. Evaluation focuses on navigation success, path efficiency, and recovery.
Reliable Multimodal Intelligence under Partial Observability
My earlier work asks how models can learn when views or modalities are missing. That foundation now supports a broader embodied question: how should an agent reason and act when observations are incomplete, uncertain, or viewpoint-dependent?