Robot learning Mobile manipulation

Map-Aware Visuomotor Policies for Mobile Manipulation

Jinhe Tang1,†, Ruixiao Dai1,†, and Weiming Zhi1*

1School of Computer Science, The University of Sydney, Australia

†These authors contributed equally.

MAVP learns explicit base-pose targets in a shared map frame and tracks them with localisation feedback, helping a mobile robot reach the right place for manipulation.

arXiv:2609.26378 Watch six task demos Explore the method
Original six-task artwork showing a mobile robot performing drawer packing, disassembly, box packing, conveyor picking, dual-drawer return, and bag packing
Six real-world tasks, from drawer packing to bag packing.

Overview

One shared spatial reference, from demonstrations to execution.

Demonstrations can disagree about where the base should go when their motion is expressed only as local velocity. MAVP reconstructs a static map from demonstrations, uses that map to align base poses, and predicts pose targets alongside arm and gripper actions.

Method

Map, localise, act.

Figure 2 artwork from the paper source: shared map construction, localisation, and policy learning.

Original Figure 2 artwork showing shared map construction, live localisation, and map-aware policy learning

Experiment videos

Six real-world mobile manipulation tasks

Each clip shows a task rollout. Footage is accelerated for viewing; people visible in the original recordings have been masked or removed from these edits. All six clips use an original instrumental MAVP soundtrack.

01

Drawer Packing

Pick up the eraser, open the drawer, place the eraser and two toys inside, and close it.

02

Disassemble and Deliver

Remove the peg, travel to the box, and place all parts inside.

03

Lidded Box Packing

Approach and open the box, place an object inside, and close the lid.

04

Conveyor Picking

Track and grasp an object moving on the conveyor.

05

Dual-Drawer Return

Return two objects to their corresponding drawers.

06

Bag Packing

Take a bag from the rack, pack the table objects, and hang the bag back.

Experimental results

Map-frame targets improve success across tasks and policies.

All results below are drawn from the paper. Each policy was evaluated over 50 physical rollouts per task; success requires completing every task step.

Table V · Main comparison

Six tasks, three policy families

V uses unanchored velocity without a pose input. P predicts map-frame pose targets with pose input and localisation feedback.

Table V. Task success rates in percent over 50 rollouts per task
TaskACTDiffusion policyFlow matching
VPVPVP
Disassemble and Deliver38%90%42%86%32%88%
Drawer Packing20%64%44%62%38%66%
Conveyor Picking42%66%38%74%36%76%
Lidded Box Packing44%88%52%56%50%60%
Bag Packing18%66%28%68%32%70%
Dual-Drawer Return22%68%30%70%36%74%

Source: paper Table V. Bold cells mark the highest success rate for each task.

Table II · Controlled variants

What each component adds

Success rates for ACT variants on two tasks, with the same evaluation protocol.

Table II. Success rates for six base-control variants
Method / variantDisassemble & DeliverLidded Box Packing
Unanchored velocity38%44%
Odometry pose16%16%
Map-frame pose without pose input74%58%
Velocity with pose input60%64%
Map-frame pose with pose input84%76%
Map-frame pose with pose input + augmentation90%88%

Pose denotes the current map-frame base-pose input; augmentation adds pose noise during training.

Table VI · Ongoing localisation

Keep the map frame aligned throughout execution

Success when alignment is fixed after the first map localisation versus updated during deployment.

Disassemble and Deliver

Initial only
64%
Ongoing
90%

Lidded Box Packing

Initial only
52%
Ongoing
88%

Source: paper Table VI. Ongoing-localisation results are reused from Table II.

Full paper

Read the complete method and experiments.

Available as arXiv:2609.26378. The tables above show the main comparisons. The PDF contains the full setup, evaluation protocol, ablations, and discussion.

Open the arXiv PDF

Citation

BibTeX

@misc{tang2026mavp,
  title         = {{MAVP}: Map-Aware Visuomotor Policies for Mobile Manipulation},
  author        = {Jinhe Tang and Ruixiao Dai and Weiming Zhi},
  year          = {2026},
  eprint        = {2609.26378},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2609.26378}
}