Contents

MoPA: Coordinated Mobile Manipulation via
Subsystem-Specific Perception Alignment

Guangyu Chen1,2,* Qiwei Liang2,3,* Shaolong Zhu1,2,* Tianxing Chen2,4,* Zikuan Xiao2 Yifan Xie1 Lingfeng Zhang1 Ping Luo4 Renjing Xu3,† Wenbo Ding1,2,†

1 THU 2 Xspark AI 3 HKUST (GZ) 4 HKU

* Equal contribution † Corresponding author

arXiv Code Coming soon Hugging Face

METHOD

MoPA architecture: Dual Perceptual Streams extract subsystem-specific representations from shared Qwen3-VL context using mutually masked Mobile and Manip Query banks. Perception2Action Adaptation jointly updates each query bank with its matched action stream, while the two action streams communicate for whole-body coordination.
Overview of MoPA. (a) A shared vision–language foundation model produces separate Mobile Query and Manip Query streams that condition dedicated Base and Manipulation experts in a joint action head. (b) Structured visibility patterns in the VLM and action head: direct attention between the query streams is masked, each query stream connects to its matched action stream, and the two action streams remain coupled for whole-body coordination.

Distinct queries
make a difference.

Three architectures: shared perception with one action head, shared perception with separate action experts, and MoPA's subsystem-specific perception with coordinated experts. Bar charts compare mean SetTable skill success and mean real-world full-task success rates.
Mobile-manipulation paradigms and performance. (a) Shared perception conditions a unified action head. (b) Separate base and manipulation experts share perceptual conditioning. (c) MoPA combines subsystem-specific perception with action-level coordination. Bottom: average success rates in simulation and real-world experiments.
Mobile Query attention and Manip Query attention visualizations shown alongside the same RGB robot observation.
Query attention visualization. (a) Mobile Query attention, (b) the RGB observation, and (c) Manip Query attention.
A cosine-similarity heatmap for Manip and Mobile queries, and a t-SNE plot in which orange Manip Query states and purple Mobile Query states occupy largely distinct regions with some overlap.
Query representations during inference. Left: cosine-similarity heatmap. Right: t-SNE projection, with orange and purple denoting Manip Query and Mobile Query states, respectively.
BENCHMARKS

Stronger across
all three suites.

ManiSkill-HAB

Mean skill success rate (%) ↑

SetTable

7 skills · Table I
  • DP3 21.6%
  • ACT 23.6%
  • DP 28.8%
  • RDT-1B 42.9%
  • AC-DiT 55.6%
  • π0 59.1%
  • AnchorVLA 64.0%
  • π0.5 68.9%
  • MobileWAM 73.0%
  • GeoHAT 79.3%
  • InCoM 83.8%
  • MoPA 88.3%

TidyHouse

18 skills · Table II
  • ACT 16.9%
  • DP 15.2%
  • DP3 30.5%
  • DSPv2 21.7%
  • InCoM 47.8%
  • GeoHAT 51.8%
  • π0.5 55.2%
  • MoPA 72.2%

PrepareGroceries

18 skills · Table II
  • ACT 14.8%
  • DP 9.1%
  • DP3 15.7%
  • DSPv2 16.9%
  • InCoM 40.5%
  • GeoHAT 40.2%
  • π0.5 60.7%
  • MoPA 67.8%
REAL WORLD

Ready for
the real world.

MoPA jointly trained on 4 tasks in one policy.
200 demos and 20 evals per task.

Real robot execution sequences for collecting fruit into a bowl, sorting blocks into color-matched cups, transferring a block between tables, and retrieving an object from a microwave, alongside the robot platform and baseline comparison chart.
Real-world evaluation of MoPA. (a) Execution sequences for fruit collection, color-matched sorting, cross-table block transfer, and microwave object retrieval. (b) Robot platform. (c) Success rates compared with DP, AC-DiT, and π0.5.

01Color-matched sorting

4× speed
MoPA
DP
AC-DiT
π0.5

02Fruit collection

4× speed
MoPA
DP
AC-DiT
π0.5

03Cross-table block transfer

4× speed
MoPA
DP
AC-DiT
π0.5

04Microwave object retrieval

4× speed
MoPA
DP
AC-DiT
π0.5
CITATION

Build on this work.

@misc{mopa,
  author = {Chen, Guangyu and Liang, Qiwei and Zhu, Shaolong and
            Chen, Tianxing and Xiao, Zikuan and Xie, Yifan and
            Zhang, Lingfeng and Luo, Ping and Xu, Renjing and
            Ding, Wenbo},
  title = {{MoPA}: Coordinated Mobile Manipulation via
           Subsystem-Specific Perception Alignment},
  note  = {Manuscript}
}