VLA Models Bridge Bimanual Manipulation and Drone Navigation

Vision Language Action (VLA) models are emerging as the dominant framework for learning-based robotics, combining visual perception, natural-language understanding, and action generation into a single foundation model. A new comprehensive review by Inkyu Sa and four co-authors examines 183 contributions spanning 2017–2026, organized across seven dimensions: VLA architectures, training recipes, action representations, bimanual coordination, UAV navigation/control, language grounding, and cross-cutting issues like memory and world models. The paper highlights how bimanual manipulation—requiring two arms with 7 degrees of freedom each to work in concert for tasks like folding or assembly—serves as the most demanding testbed for VLA systems, and structurally parallels the challenges faced by drones coordinating thrust, attitude, and gripper commands under strict latency and payload constraints.

The review's key finding is that the coordination strategies, training recipes, and action representations developed for bimanual VLAs are transferable to unmanned aerial systems. The authors identify 14 specific research directions that apply to both domains, potentially unifying robot manipulation and autonomous drone navigation under a single VLA framework. This 56-page paper, with 11 figures and 16 tables, provides a roadmap for future development of VLA models across robotics and aerial systems, offering implications for everything from industrial automation to delivery drones and search-and-rescue operations.