The new VAMOS method divides robot navigation into two levels: general-purpose planning and implementation that considers the body's capabilities are separated. The goal is for the same high-level planner to generalize across various environments, but the physical constraints of an individual robot—such as whether the device can climb stairs—guide the final route.

The approach is based on a hierarchical vision-language-action model (VLA), where semantic planning is detached from the so-called embodiment anchoring. The general planner learns from open-world, diverse data, while the specialized affordance model learns the robot's own capabilities and safety-related constraints affordably in simulation. Affordances here mean what actions the environment and the robot together enable or prevent.

Technically new is the interface where the high-level planner proposes multiple route options directly in the image plane. The affordance model evaluates these proposals in terms of the robot's capabilities and rearranges them, selecting the path that not only leads to the goal but is also feasible for the specific device. This way, semantic understanding of the environment and concrete mobility are made to work together without either level overly constraining the other.

According to the authors, the solution is trained by dividing learning into context: the planner utilizes broad, heterogeneous data, while affordances are learned safely and cost-effectively in a simulation that reflects the specific robot's physics. The work also reports real-world experiments where the VAMOS approach is practically evaluated.

If the method becomes widespread, the same planning component could guide very different platforms, and each robot's own capability limits would be considered with a lightly interchangeable affordance layer. This promises modularity and steerability for navigation tasks where the range of environments and device constraints vary.

Source: VAMOS: A Hierarchical Vision-Language-Action Model for Capability-Modulated and Steerable Navigation, ArXiv (AI).

No news available