ViT Quadrotor Obstacle Avoidance
📝 Detailed writeup coming soon — key information is shown below.
ViT-based Quadrotor End-to-End Obstacle Avoidance Thesis
Overview
Traditional UAV controllers suffer from perception-to-control latency in high-speed agile flight, especially when relying on classical pipelines. This thesis implements a Vision Transformer (ViT) based end-to-end control framework for quadrotor obstacle avoidance in forest environments, comparing it against CNN and LSTM-based baselines.
Baseline Reproduction
- Reproduced the ViT and ViT-LSTM baselines from an ICRA 2025 paper within the Flightmare + Dodgelib simulation environment
- Ran comparative experiments confirming ViT's stronger out-of-distribution generalization compared to CNN-based controllers in unseen forest environments
Input Modality Extension
- Rebuilt the expert data collection pipeline in Flightmare, recording RGB frames, drone pose, and velocity commands
- Modified the model input from single-frame depth maps to multi-frame RGB image sequences, allowing temporal context
- Added an IMU parameter fusion channel to the ViT encoder to incorporate inertial information
Results
- Achieved 73.26% obstacle avoidance success rate in forest simulation (Flightmare + Dodgelib)
- Validated closed-loop control on a local Ubuntu 20.04 dual-boot setup
Training curves, ablation results, and simulation recordings will be added here.
📝 详细内容即将补充，以下为主要信息。
基于 ViT 的四旋翼无人机端到端避障系统 毕设
项目概述
传统无人机控制器在高速敏捷飞行中存在感知到控制的延迟问题，尤其是在依赖经典管线的情况下。本毕设基于 Vision Transformer（ViT）实现四旋翼无人机端到端避障控制框架，并在树林仿真环境中与 CNN 和 LSTM 基线进行比较。
基线复现
- 在 Flightmare + Dodgelib 仿真环境中复现 ICRA 2025 论文的 ViT 与 ViT-LSTM 基线
- 对比实验确认 ViT 在陌生树林环境中相比 CNN 控制器具有更强的分布外泛化性
输入模态扩展
- 基于 Flightmare 重构专家数据采集流程，记录 RGB 图像、无人机位姿与速度指令
- 将模型输入从单帧深度图改为多帧 RGB 图像序列，引入时序上下文
- 在 ViT 编码器中增加 IMU 参数融合通道以引入惯性信息
结果
- 在树林仿真环境（Flightmare + Dodgelib）中达到 73.26% 避障成功率
- 在本地 Ubuntu 20.04 双系统上完成闭环控制验证
训练曲线、消融实验结果与仿真录像将在后续补充。