MobileVLA-R1 2.0: RL-Enhanced Reasoning for Mobile Robot Control

MobileVLA-R1 2.0. RL-Enhanced Reasoning for Mobile Robot Control. By Ting Huang and four others, from Peking University, the South China University of Technology, and one other institution. Published on the fifth of September. Its code repository is public and real: about twenty source files.

For vision-language-action systems on mobile robots, turning a natural-language instruction into a reliable, executable action is still a basic difficulty. The gap sits between high-level semantic reasoning and low-level control of locomotion and manipulation. Existing approaches often either reason implicitly, or predict actions in one monolithic step. Either way, it is hard to keep long-horizon decisions coherent while still producing actions that are precise and adaptable. A robot told to walk to the kitchen and pick something up has to hold the plan together over many steps, and also move its legs and arm correctly at each one. Doing both inside one undivided prediction is where the existing approaches struggle.

MobileVLA-R1 2.0 is a VLA framework enhanced with reinforcement learning, which explicitly couples structured embodied reasoning to executable robot control. It learns reasoning at multiple granularities over embodied trajectories, drawing on two signals: supervised Chain-of-Thought alignment, and reinforcement learning. The aim is consistency between what the model reasons and what it then does, beyond what purely behavioral supervision would give. Behavioral supervision alone teaches a model to copy actions. Adding explicit reasoning, and then reinforcing it, is meant to make the stated reasoning and the action that follows it agree.

On the action side, a reasoning-conditioned action decoder maps multimodal reasoning representations to task-level action targets. Robot controllers then translate those targets into commands specific to each embodiment. That split is what lets one framework cover both locomotion and manipulation. Perception, reasoning and action share a unified interface, while high-level action generation stays decoupled from the actuation details of any particular robot. The model decides what should happen at the level of the task. How a given body carries that out is left to the controller built for that body.

Evaluation covers language-guided navigation, quadruped control, and humanoid mobile manipulation. That includes VLN-CE, QUARD, and real-world deployments on Unitree Go2 and G1 robots. MobileVLA-R1 2.0 consistently outperforms strong VLA baselines. Measured against the original MobileVLA-R1, it adds one point six points of success rate on average on VLN-CE, and ten points of full-task success on real-world G1 mobile manipulation tasks. It also shows robust long-horizon instruction following and closed-loop execution across different robotic platforms.

The design turns on where the boundary is drawn. Reasoning and task-level targets live in one shared model, and everything specific to a robot's body is handed to that robot's own controller. That division is what lets the same reasoning drive a four-legged robot and a humanoid, and it is also where the reinforcement learning earns its place: keeping what the model says it will do tied to what it then does.