# ASMA: An Adaptive Safety Margin Algorithm for Vision-Language Drone Navigation via Scene-Aware Control Barrier Functions

**Authors:** Sourav Sanyal, Kaushik Roy  
**arXiv:** [2409.10283](https://arxiv.org/abs/2409.10283)  
**Date:** 2024-09-16  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2409.10283v1/

## Key points

- Control barrier functions (CBFs) are essential in ensuring safety in vision-language navigation (VLN)
- The study "ASMA" by Sourav Sanyal and Kaushik Roy focuses on enhancing safety mechanisms for a teleoperated drone in a VLN setting
- A novel scene-aware CBF formulation incorporating ego-centric observations through an RGB-D sensor is introduced to improve safety features
- The research implements a vision-language understanding module using the CLIP model for specifying landmarks and verifying them with the YOLO object detector before navigation commands
- ASMA, an Adaptive Safety Margin Algorithm, dynamically adjusts the drone's depth map to track moving objects in real-time, enabling adaptive responses to unpredictable conditions
- Experimental validation of ASMA on a Parrot Bebop2 quadrotor within the Gazebo simulation environment shows significant success rate improvements compared to baseline approaches without CBFs
- ASMA demonstrates efficient recovery from unsafe situations, showcasing its effectiveness in enhancing safety measures for VLN-powered drones

## Summary

In the rapidly evolving field of vision-language navigation (VLN), ensuring robust safety mechanisms remains a critical challenge. Control barrier functions (CBFs) have emerged as efficient tools to guarantee safety by solving optimal control problems. In this study titled "ASMA: An Adaptive Safety Margin Algorithm for Vision-Language Drone Navigation via Scene-Aware Control Barrier Functions," authors Sourav Sanyal and Kaushik Roy address this challenge by focusing on a teleoperated drone operating in a VLN setting. The researchers introduce a novel scene-aware CBF formulation that incorporates ego-centric observations obtained through an RGB-D sensor to enhance safety features. To establish a baseline, they implement a vision-language understanding module utilizing the contrastive language image pretraining (CLIP) model. This module allows users to specify landmarks in natural language, which are then verified using the YOLO object detector before triggering navigation commands. To further improve safety during navigation, the authors propose ASMA, an Adaptive Safety Margin Algorithm. This algorithm dynamically adjusts the drone's depth map to track moving objects in real-time, enabling scene-aware CBF evaluation on-the-fly. By identifying potential risks within the environment, ASMA facilitates adaptive responses to unpredictable conditions, thereby ensuring optimal safety boundaries for drone actions powered by VLN technology. The experimental validation of ASMA was conducted using the robot operating system (ROS) middleware on a Parrot Bebop2 quadrotor within the Gazebo simulation environment. Results demonstrate that ASMA significantly enhances success rates by 59.4% - 61.8% compared to baseline approaches lacking CBFs, with only minor increases of 5.4% - 8.2% in trajectory lengths. Additionally, ASMA showcases its capability to recover from unsafe situations efficiently, highlighting its effectiveness in enhancing safety measures for VLN-powered drones. Overall, this research contributes valuable insights into advancing safety mechanisms in VLN applications through innovative algorithms like ASMA, paving the way for safer and more reliable autonomous drone navigation systems in complex environments.