A small wheeled robot and a larger robot with a different steering system cannot move through a narrow corridor in exactly the same way. The command to turn right produces different motions. A team led by Zhidong Wang proposed teaching one program to plan paths for several kinds of robots while still respecting those differences.
The method uses a world model[1]: software that predicts how the scene around a robot may change after it moves. But pictures alone are not enough. The researchers also describe how each particular robot accelerates and turns. The planner can therefore consider what the machine can actually do, not merely which space looks empty.
In computer experiments, the researchers used seven robots representing five kinds of movement. The shared program succeeded 3.8 to 10.6 percentage points[2] more often than some comparison methods. They then drove a real robot manually for three minutes to learn how it moved, before testing navigation in an unfamiliar indoor space.
The method couples a shared world model with a differentiable reduced-order model of each robot's motion through the change in camera viewpoint. The world predictor learns from multiple platforms, while kinematic parameters are identified separately. During imagined rollouts, a proposed command first passes through the robot template and then the visual world predictor. This makes the planner sensitive to turning and speed limits without training an entirely separate policy for each platform.
That real-world test matters, but it does not show that the same program works for every vehicle or on staircases. The study specifically discusses movement on flat ground and particular robot designs.
The problem is not just whether a robot sees an empty passage. Different machines steer, brake and accelerate differently. A planner looking only at the room may propose a turn the machine cannot make. The researchers added a compact description of each robot's movement to the shared world model. The result combines common knowledge of the scene with physical constraints specific to each machine, rather than sending identical commands to every robot.
In the physical experiment, the researchers identified a new robot's parameters from three minutes of teleoperated driving and then tested unfamiliar indoor navigation. The flat-ground assumption does not cover stairs, steep slopes or contact-heavy manoeuvres. The numerical baseline gains were measured in simulation; the physical test supports a narrower transfer claim.
A real robot needed three minutes of manual driving to identify its movement characteristics. That is neither a complete installation time nor a test of stair climbing. The experiment raises a practical question about how little experience with a new machine a shared planner needs before its predictions become useful.