A mobile robot completed more multi-step tasks when its learned controller aimed for positions on a shared map rather than predicting base velocities, researchers at the University of Sydney report. The map gave the robot a way to check where it was and correct its route before a positioning error disrupted the next grasp. In lab tests, the map-aware system beat the velocity-predicting baseline across six tasks and three kinds of learned policy.
For a robot that moves between workstations, reaching roughly the right area is not enough. Its base position affects what its arms can reach and what its cameras can see. A policy trained from human-guided demonstrations can accumulate small movement errors until it arrives at the wrong angle for the next step.
A shared map turns movement into a position target
The researchers' Map-Aware Visuomotor Policies system builds a static map from demonstration footage, leaving out the robot's moving arms and other changing objects. It records demonstrated routes in that same coordinate system. During a task, camera observations locate the robot on the map; the policy predicts a destination for its wheeled base alongside arm and gripper actions. A lower-level controller then steers toward that destination and uses the difference between target and current position to correct drift. Training also adds small errors to the position input so the policy can tolerate imperfect localisation.
Map-aware system completed more lab tasks
For one of the three policy families, the complete map-aware setup finished a disassembly-and-delivery task in 90% of 50 trials, compared with 38% for a baseline that predicted base velocities without a map-position input. It finished lidded-box packing in 88% versus 44%. The researchers tested six physical tasks, each with 50 demonstrations and 50 evaluation trials per policy, and found higher completion rates for the map-aware setup in all six tasks across three policy families. Success meant finishing every step, not merely reaching an object.
That broad comparison changes several parts of the system together: the map-based position input, destination targets, feedback and training for localisation noise. In a narrower two-task comparison that gave both policies the same position input and omitted noise augmentation, targeting positions still beat predicting velocities: 84% versus 60% on disassembly and delivery, and 76% versus 64% on lidded-box packing. Keeping map localisation active through a task also improved completion over aligning only at the start in a separate two-task test.
The tests used one robot in mapped workspaces whose basic scene geometry stayed fixed.
Sources
- MAVP: Map-Aware Visuomotor Policies for Mobile Manipulation, University of Sydney preprint, 22 September 2026; full paper with methods and Tables II, V and VI.
- MAVP research project and task videos, by the paper's authors.
- Lead illustration: Figure 2 from the authors' paper, CC BY 4.0, used without modification.