HKUST team proposes RAVEN, an active SLAM framework for quadrotors

Xinyi Chen and other authors propose RAVEN, a unified active visual SLAM framework for quadrotors that places exploration planning and active loop closure within the same framework and explicitly accounts for localization uncertainty. The framework first identifies frontiers and active loop-closure candidates and generates a global tour plan, then evaluates them with a utility vector, and finally a perception-aware navigation planner generates smooth trajectories based on an online Fisher information field to reduce visual state estimation error; the authors state that a priority parameter from 0 to 1 can be used to adjust the emphasis between exploration efficiency and localization accuracy. Subsequent reporting will state that the team is identified as the HKUST Shen Shaojie team, and that the parameter λ (0–1) determines whether the bias is toward exploration or localization, with no manual weight tuning needed throughout. The report states that fully onboard flights were conducted in three scenarios—Flight Arena with OptiTrack ground truth, the all-white-wall Campus Hallway, and an underground Parking Lot—with 3–5 test flights for each of three λ settings; in the two scenarios without ground-truth trajectories, evaluation was done indirectly using the nearest-neighbor distance from the reconstructed map to the ground-truth point cloud, and λ=1 reduced the error by 50% and 62.5%, respectively, relative to λ=0. The authors state that fully onboard field experiments were completed with a custom stereo-inertial quadrotor, verifying its practicality in challenging environments; the paper has been conditionally accepted by IEEE Transactions on Robotics, and the source code will be released to the community. These are all statements by the authors and have not yet been independently verified.

AI 根据报道生成 · 5 days ago更新