Flow-Aided Flight Through Dynamic Clutters From Point To Motion

TL;DR

The paper tackles autonomous flight through highly dynamic clutter by eliminating object detection and prediction modules, relying instead on a fixed-shape depth-distance map and multi-frame point flow to encode environmental changes. A PPO-based RL policy consumes these representations to learn early, motion-aware avoidance via a relative-motion-guided distance field, achieving robust sim-to-real transfer on quadrotors. Key contributions include the distance-map encoding, pre-trained point flow for motion features, and a lightweight end-to-end framework that outperforms baselines in simulation while maintaining safe real-world performance. The work demonstrates that end-to-end perception-to-action systems can achieve reliable navigation in dynamic clutters with reduced sensing and computation requirements.

Abstract

Challenges in traversing dynamic clutters lie mainly in the efficient perception of the environmental dynamics and the generation of evasive behaviors considering obstacle movement. Previous solutions have made progress in explicitly modeling the dynamic obstacle motion for avoidance, but this key dependency of decision-making is time-consuming and unreliable in highly dynamic scenarios with occlusions. On the contrary, without introducing object detection, tracking, and prediction, we empower the reinforcement learning (RL) with single LiDAR sensing to realize an autonomous flight system directly from point to motion. For exteroception, a depth sensing distance map achieving fixed-shape, low-resolution, and detail-safe is encoded from raw point clouds, and an environment change sensing point flow is adopted as motion features extracted from multi-frame observations. These two are integrated into a lightweight and easy-to-learn representation of complex dynamic environments. For action generation, the behavior of avoiding dynamic threats in advance is implicitly driven by the proposed change-aware sensing representation, where the policy optimization is indicated by the relative motion modulated distance field. With the deployment-friendly sensing simulation and dynamics model-free acceleration control, the proposed system shows a superior success rate and adaptability to alternatives, and the policy derived from the simulator can drive a real-world quadrotor with safe maneuvers.

Figures (8)

• Figure 1: Outdoor flight in natural clutters with pedestrians. The left part shows the executed trajectory of the neural controller. The snapshots in the right part demonstrate the avoidance response when confronted with both dynamic pedestrians and static trees. The movement of the pedestrians can be seen clearly in the “ghost trail” of the map built after the experiment.
• Figure 2: Illustration of the change-aware LiDAR sensing representation. The distance map is encoded by robot-centric omnidirectional depth perception, which ensures safe sensing of obstacle details through nearest selection. The point flow is derived from pseudo-RGB image pairs stacked with 3 consecutive grayscale images, which are fed into a flow estimation module that does not update during RL training for motion feature extraction.
• Figure 3: Implementation of the autonomous flight system from point to motion. State estimation xu2022fast of the quadrotor is realized through the collaborative use of the LiDAR point clouds and the IMU data, which is a robust real-world substitute for ground truth states used in training. The unit vector pointing to the goal $\boldsymbol{g}$, the velocity $\boldsymbol{v}$ and the last output $\boldsymbol{a}_{\text{last}}$ form the state inputs of the system, which are used to feed into the MLP fusion module together with the LiDAR sensing representation features extracted by the CNN encoder. The output of the system is the desired acceleration command $\boldsymbol{a}$.
• Figure 4: Effects of reshaping the distance field of dynamic obstacles. The distance field is shown in a range of $5 \times 5\ m^2$ when the dynamic obstacle is at the same orthogonal velocity of $1\ m/s$ as the quadrotor. This exhibits the extra threat given to $\theta\in (0, \pi/2)$ is roughly symmetrical about the relative velocity $\boldsymbol{v}_{\text{dobs}} - \boldsymbol{v}$ and skewed toward the absolute obstacle velocity $\boldsymbol{v}_{\text{dobs}}$.
• Figure 5: Training Environment and Obstacle Setup. (a) Parallel training in 128 environments with centrosymmetrically set start-goal points, and walls are set at the same height as the start-goal points. (b) Top-down view of the trajectories of the dynamic obstacles performing uniform linear motions.