I'm curious about the underlying methods drones use to navigate around moving obstacles without human input. Specifically, how do perception algorithms combine data from lidar, camera, or radar to build a real‑time map, and what role do planning frameworks like RRT* or neural‑network policies play in generating safe trajectories? Also, how is safety guaranteed when sensor noise or unexpected objects appear? Would love to hear explanations or references on current best practices.
How do autonomous drones handle obstacle avoidance in dynamic environments?
👁️ 30 görüntüleme💬 2 cevap❤️ 0 beğeni
2 Cevap
Kanka, geçen hafta bir mini‑drone’ı ROS ile denediğimde lidar ve tek bir kamera verisini EKF ile birleştirip anlık nokta‑bulut haritası çıkardık, sonra da RRT*‑ı basit bir safety‑margin ekleyerek çalıştırdık; sensör gürültüsü olduğunda algoritma bir adım geri çekilip yeni bir yol buluyor, bu da çarpışmayı önlüyor. Bence, dinamik ortamda ekstra bir “neural‑policy” katmanı eklemek, ani hareketli nesneler geldiğinde daha yumuşak kaçış manevraları sağlıyor, çünkü ağ anlık olarak “risk haritası” üretip RRT*’a yön veriyor.
I’ve been tinkering with a DIY quad‑copter for the past year, and the biggest headache turned out to be keeping it clear of people and pets moving around the living room. I ended up fusing a 2‑D lidar slice (Hokuyo UST‑10LX) with a forward‑facing RGB‑D camera using an EKF‑based sensor‑fusion node in ROS 2; the lidar gives precise range updates at 10 Hz while the camera supplies semantic masks (person, pet, chair) via a tiny MobileNet‑SSD model running on the Jetson Nano. The fused data populates a 3‑D occupancy grid that’s updated every 100 ms, and I feed that grid into a local RRT* planner that replans whenever a new obstacle appears within a 2‑m horizon. To keep the planner fast, I prune the tree after each iteration and bias the sampling towards the current goal direction, which gives smooth, collision‑free waypoints even when the couch is being dragged across the floor.
When the sensor confidence drops—say the lidar gets a bit noisy from sunlight or the camera’s depth stream glitches—I switch to a reactive neural‑network policy that was trained on simulated cluttered rooms (using AirSim). The network takes the same fused occupancy slice and outputs a velocity command that pushes the drone sideways while maintaining altitude, essentially acting as a safety net until the next RRT* pass. I also run a watchdog that monitors the covariance of the EKF; if it spikes, the drone automatically hovers and lands. In practice this combo of deterministic planning and learned fallback has kept my indoor flights safe despite the occasional stray cat or moving vacuum cleaner.