MoPA: Coordinated Mobile Manipulation via Subsystem-Specific Perception Alignment

• Prior work expands perceptual context and specializes action computation, yet heterogeneous action branches generally remain conditioned on shared or globally fused representations.Source: §I · p. 1
• Mobile manipulation couples base and arm motion despite their different perceptual needs, while existing shared representations leave subsystem-specific perception–action correspondence implicit.Separating action heads alone does not guarantee distinct perceptual conditioning, and shared projection or fusion layers can receive updates from both branches.Source: §I · pp. 1–2
• MoPA’s isolated Mobile Query and Manip Query banks extract subsystem-relevant evidence from a shared vision–language context, while matched query–action updates and coupled action streams coordinate generation.The structured decoder lets each query bank interact only with its matched action stream while the two action streams exchange evolving control states.Source: §I · p. 2
• MoPA achieves the highest macro-averaged skill success rates among compared methods on SetTable (88.3%), TidyHouse (72.2%), and PrepareGroceries (67.8%), exceeding each suite’s strongest baseline by 4.5, 17.0, and 7.1 percentage points, respectively.Source: §I · p. 2

II. RELATED WORK

• Mobile-manipulation and vision–language–action methods improve grounding or specialize control, but typically select subsystem-specific evidence within the action decoder or retain a shared action interface.Source: §II · p. 2
• World–action methods specialize through future latents or action experts, whereas MoPA establishes an explicit persistent interface between shared perceptual context and coordinated action streams.Source: §II · p. 2

III. METHOD

• The method derives branch-specific perceptual representations from shared vision–language tokens, adapts them through Perception2Action Adaptation, and coordinates the two action streams with coupled conditional flow matching.Fig. 2 depicts masked direct attention between query banks, matched query-to-action connections, and continued coupling between action streams.Source: §III · p. 2, Figure 2 · p. 3

• At each time step, the policy receives head- and wrist-camera RGB images, proprioceptive state, and a language instruction, and predicts a whole-body action chunk of horizon H.Source: A. Problem Formulation · p. 3
• The manipulation and mobility branches control the arm and gripper, and the torso and base, respectively; joint modeling learns their coordination from whole-body demonstrations.Source: A. Problem Formulation · p. 3

B. Method Overview

• MoPA encodes head- and wrist-view images with the instruction, then uses separate query banks and branch-specific proprioception to condition action generation.A structured Mixture-of-Transformers decoder pairs each perceptual stream with its corresponding action stream, and coupled conditional flow matching generates both chunks synchronously.Source: B. Method Overview · p. 3

C. Dual Perceptual Streams

• Both query banks attend to the shared vision-language tokens but cannot attend directly to each other, producing distinct representations routed only to their matched action streams.Separate aggregation paths and supervision from the corresponding action branches encourage each bank to retain task-relevant evidence.Source: C. Dual Perceptual Streams · p. 3

D. Perception2Action Adaptation

• Perception2Action Adaptation uses separate parameters for each action branch while coupling their evolving states, so each query stream interacts with its matched action stream and the action streams exchange information.Query states update jointly with action tokens at every decoder layer; final action tokens produce manipulation and mobility flow velocities.Source: D. Perception2Action Adaptation · pp. 3–4

E. Coupled Conditional Flow Matching

• Demonstration chunks for both branches come from the same interval and share a flow time, preserving temporal correspondence across the action horizon.Each clean chunk is paired with an independent Gaussian noise chunk, and the shared flow time defines interpolation from noise to actions.Source: E. Coupled Conditional Flow Matching · p. 4
• The joint action head predicts both branch velocities from projected query states and encoded branch sequences, with each prediction able to depend on the other branch's current state.The predicted field depends on interpolated actions, perceptual conditioning, and flow time.Source: E. Coupled Conditional Flow Matching · p. 4
• The branch-normalized flow-matching objective prevents the higher-dimensional branch from dominating solely because it has more action channels.The squared error is normalized by each branch's H × d_k entries, with λ_k balancing branch contributions.Source: E. Coupled Conditional Flow Matching · p. 4
• At inference, both chunks start from independent Gaussian noise and are integrated synchronously with N Euler steps, then concatenated into a whole-body chunk before execution and replanning.Both velocities are evaluated from the same pre-update pair of chunks at each step.Source: E. Coupled Conditional Flow Matching · p. 4

IV. EXPERIMENTS

• The evaluation includes simulation benchmarks.The section identifies these as Sec. IV-A.Source: §IV · p. 4
• Ablations examine MoPA's core components, while further analyses explore query capacity and representational differentiation.These analyses are identified as Secs. IV-B and IV-C.Source: §IV · p. 4
• The evaluation also assesses real-world performance.This assessment is identified as Sec. IV-D.Source: §IV · p. 4

Simulation results

• 88.3% on SetTable, 72.2% on TidyHouse, and 67.8% on PrepareGroceries are the highest macro-averaged skill success rates among compared methods.These results exceed the strongest baseline on each suite by 4.5, 17.0, and 7.1 percentage points, respectively.Source: Table I · p. 5, Table II · p. 5, A. Simulation Experiments · p. 5

Skill-level results

• Across SetTable and TidyHouse, the reported improvement is larger on TidyHouse’s Pick than Place category (22.5 vs. 11.5 percentage points), while InCoM retains the lead on SetTable’s Place Apple and TidyHouse’s Place category.Source: A. Simulation Experiments · p. 5
• MobileWAM reaches 99.3% on Open Fridge and 100.0% on Close Drawer, but scores 46.0% on each picking task, versus MoPA’s 84.3% on Pick Apple and 89.0% on Pick Bowl.The passage presents this contrast as motivation for subsystem-aligned perceptual conditioning alongside action specialization.Source: A. Simulation Experiments · p. 5

Analysis

• Mobile Query attends to floor and scene structure, while Manip Query focuses on the shelf around the target object, consistent with their intended roles.Performance drops under Shared Query, w/o Joint Attention, and w/o Corresponding Query Access support subsystem-specific conditioning and adaptation during decoding.Source: A. Simulation Experiments · p. 5

Ablation setup

• The ablation variants use the same backbone and training and evaluation protocols, while retaining both action experts and interaction between their streams.The comparisons examine perceptual representation construction and how representations are updated and matched to action branches.Source: B. Ablation Studies · p. 6

Perceptual conditioning

• Separating subsystem queries raises success to 88.3%, compared with 74.7% for Shared Query, without changing total learned query capacity.Direct VLM Conditioning replaces learned queries; the comparisons support learnable extraction and subsystem-specific conditioning, while query aggregation alone does not explain the overall improvement.Source: Table III · p. 6, B. Ablation Studies · p. 6

Joint query–action adaptation

• 70.3% success without Joint Attention is 18.0 percentage points below the full model, although query parameters remain trainable and matched query access is preserved.This variant fixes query hidden states within each action-head evaluation; the result supports adapting perceptual representations to evolving actions rather than supplying static summaries.Source: Table III · p. 6, B. Ablation Studies · p. 6

Subsystem-aligned access

• 66.3% success without Corresponding Query Access is 22.0 percentage points below the full model, despite retaining both query banks and joint updates.This variant lets each action stream read both banks, supporting explicit query–action correspondence beyond separate query aggregation and joint updates.Source: Table III · p. 6, B. Ablation Studies · p. 6

Exploratory design

• The query-bank-size sweep tests capacity sensitivity, whereas Shared Query tests subsystem differentiation at a fixed learned-query budget.The paper distinguishes these as separate experimental questions.Source: C. Exploratory Experiments · p. 7

RQ1: Effect of Query Bank Size

• Average success increases from 74.6% at q = 4 to 88.3% at q = 8, then decreases to 82.5% at q = 16; q = 8 achieves the highest success rate among tested settings.Source: C. Exploratory Experiments · p. 7

RQ2: Query Representation Differentiation

• Many cross-type query pairs have low or negative cosine similarity, and t-SNE shows predominantly distinct regions with partial overlap for Manip Query and Mobile Query states.The visualizations use query states from inference on a 20% subset of the dataset and provide qualitative evidence of differentiated representations.Source: Figure 5 · p. 7, C. Exploratory Experiments · p. 7

Results

• 76.3% average full-task success exceeds the strongest baseline, π0.5, by 12.5 percentage points across four real-world tasks.MoPA scores 85.0% on Fruit Collection, 90.0% on Color-Matched Sorting, 75.0% on Cross-Table Fruit Transfer, and 55.0% on Microwave Object Retrieval; it leads the first three and ties π0.5 on microwave retrieval.Source: D. Real-World Experiments · p. 7

Tasks and evaluation

• The tasks combine repeated pick-and-place, color-based object–receptacle matching, cross-workspace coordination, or articulated-object interaction.The evaluation uses 200 demonstrations per task and reports full-task success rates over 20 trials per task.Source: D. Real-World Experiments · p. 7

Analysis

• Dual Perceptual Streams provide separate conditions for mobility and manipulation, while Perception2Action Adaptation refines them alongside coordinated action generation.The reported success rates on fruit collection, color-matched sorting, and cross-table transfer demonstrate combined fine-grained interaction and cross-workspace execution within a single multi-task policy.Source: D. Real-World Experiments · p. 7

V. CONCLUSION

• MoPA achieves higher average success rates than evaluated baselines across three ManiSkill-HAB suites and four real-world tasks.The strongest evaluated baseline ties MoPA on microwave retrieval.Source: §V · p. 7
• MoPA links distinct mobility and manipulation perceptual representations to their action streams while preserving communication between those streams.Ablations and analyses support the complementary roles of separate perceptual conditioning, action-dependent query updates, and matched query access.Source: §V · p. 7
• Performance remains task-dependent, with a tie on microwave retrieval motivating further evaluation across more diverse environments and interaction demands.Source: §V · p. 7