TIC-VLA: A Think-in-Control Vision-Language-Action Model for Robot Navigation in Dynamic Environments

Robots in dynamic, human-centric environments must follow language instructions while maintaining real-time reactive control. Vision-language-action (VLA) models offer a promising framework, but they assume temporally aligned reasoning and control, despite semantic inference being inherently delayed relative to real-time action. We introduce Think-in-Control (TIC)-VLA, a latency-aware framework that explicitly models delayed semantic reasoning during action generation. TIC-VLA defines a delayed semantic-control interface that conditions action generation on delayed vision-language semantic states and explicit latency metadata, in addition to current observations. We further propose a latency-consistent training pipeline that injects reasoning inference delays during imitation learning and online reinforcement learning, aligning training with asynchronous deployment. To support realistic evaluation, we present DynaNav, a physics-accurate, photo-realistic simulation suite for language-guided navigation in dynamic environments. Extensive experiments in simulation and on a real robot show that TIC-VLA consistently outperforms prior VLA models while maintaining robust real-time control under multi-second reasoning latency. Project website: https://ucla-mobility.github.io/TIC-VLA/

Robots that move through busy places, such as hallways, sidewalks, warehouses, and offices, must understand human instructions while reacting immediately to people and obstacles. A major problem is that the "thinking" part of today’s robot AI can take seconds, but the robot must keep moving safely many times per second, so its understanding may already be out of date. TIC-VLA solves this by separating slow vision-language reasoning from fast motion control, while telling the controller exactly how old the reasoning is and how the robot has moved since then. The controller then uses the current camera view, the robot’s state, and this delayed guidance to choose safe short-term movements without stopping to wait for the reasoning model. The system is trained with artificial and realistic reasoning delays, first by learning from demonstrations and then by practicing in simulation. The paper also introduces DynaNav, a realistic benchmark with moving people, physical robot interactions, and indoor and outdoor navigation tasks. In simulation and real-world tests on a legged robot, TIC-VLA follows language instructions more reliably and collides less often than prior robot navigation methods, even when reasoning takes several seconds. This matters because it brings language-guided robots closer to practical use in shared human spaces, where both understanding instructions and reacting quickly are essential.