2026

1. Preprint

   AION: Aerial Indoor Object-Goal Navigation Using Dual-Policy Reinforcement Learning

   Zichen Yan, Yuchen Hou, Shenao Wang, and 3 more authors

   arXiv preprint, 2026

   Abs arXiv HTML

   We present AION, a zero-shot aerial indoor object-goal navigation system using dual-policy reinforcement learning. The system autonomously switches between exploration and target-driven navigation modes based on target visibility, leveraging RGB-D sensor input and CLIP-based multimodal features. The navigation policy is trained from scratch to master a discrete action space including take-off, navigation, and landing, and is deployed on autonomous drones using ROS 2 for real-time control.

2. Preprint

   VLA Language Understanding Benchmark: Semantic Perturbation Evaluation

   Yuchen Hou

   Under Review, 2026

   Abs

   We discover that state-of-the-art Vision-Language-Action models achieve high benchmark scores but ignore language instructions, revealing a gap between performance and true understanding. We design a systematic semantic perturbation evaluation framework to expose language grounding failures, and propose a multi-task same-scene training approach with an augmented dataset for fine-tuning.