Fanxing Li

Fanxing Li, Shengyang Wang, Fangyu Sun et al.

Existing FPV object tracking methods heavily rely on handcrafted modular pipelines, which incur high onboard computation and cumulative errors. While learning-based approaches have mitigated computational delays, most still generate only high-level trajectories (position and yaw). This loose coupling with a separate controller sacrifices precise attitude control; consequently, even if target is localized precisely, accurate target estimation does not ensure that the body-fixed camera is consistently oriented toward the target, it still probably degrades and loses target when tracking high-maneuvering target. To address these challenges, we present StableTracker, a learning-based control policy that enables quadrotors to robustly follow a moving target from arbitrary viewpoints. The policy is trained using backpropagation-through-time via differentiable simulation, allowing the quadrotor to keep a fixed relative distance while maintaining the target at the center of the visual field in both horizontal and vertical directions, thereby functioning as an autonomous aerial camera. We compare StableTracker against state-of-the-art traditional algorithms and learning baselines. Simulation results demonstrate superior accuracy, stability, and generalization across varying safe distances, trajectories, and target velocities. Furthermore, real-world experiments on a quadrotor with an onboard computer validate the practicality of the proposed approach.

Fanxing Li, Fangyu Sun, Tianbao Zhang et al.

Quadrotor control policies can be trained with high performance using the exact gradients of the rewards to directly optimize policy parameters via backpropagation-through-time (BPTT). However, designing a fully differentiable reward architecture is often challenging. Partially differentiable rewards will result in biased gradient propagation that degrades training performance. To overcome this limitation, we propose Amended Backpropagation-through-Time (ABPT), a novel approach that mitigates gradient bias while preserving the training efficiency of BPTT. ABPT combines 0-step and N-step returns, effectively reducing the bias by leveraging value gradients from the learned Q-value function. Additionally, it adopts entropy regularization and state initialization mechanisms to encourage exploration during training. We evaluate ABPT on four representative quadrotor flight tasks \li{in both real world and simulation}. Experimental results demonstrate that ABPT converges significantly faster and achieves higher ultimate rewards than existing learning algorithms, particularly in tasks involving partially differentiable rewards. The code will be released at http://github.com/Fanxing-LI/ABPT.