VLA Models Advance Robotics: Drones and Bimanual Manipulation

Vision Language Action (VLA) models represent a significant advancement in robotics, integrating visual perception, natural language comprehension, and action generation into a single foundational AI model. This allows robots to interpret high-level instructions, such as "fold the towel" or "fly to the red building," directly from visual input. VLAs benefit from extensive pre-training on internet-scale data, endowing them with broad world knowledge. The review focuses on two particularly demanding applications: bimanual manipulation and unmanned aerial robotics. Bimanual manipulation, involving two multi-degree-of-freedom arms, serves as a rigorous testbed for coordination. Similarly, drones require precise coordination of thrust, attitude, and increasingly gripper commands based on visual observations, all under strict latency and payload constraints. Analyzing 183 contributions from 2017-2026, the review organizes findings across VLA architectures, training methods, action representations, coordination strategies, and language grounding. It reveals that coordination strategies and training recipes developed for bimanual VLAs are transferable to unmanned aerial systems. The paper concludes by outlining fourteen key research directions applicable to both domains, emphasizing the shared challenges and potential for cross-pollination of solutions.