pith. sign in

arxiv: 2603.02845 · v5 · pith:ZJQF7FMTnew · submitted 2026-03-03 · 💻 cs.RO · cs.AI

SPARC: Spatial-Aware Path Planning via Attentive Agent Communication

Sayang Mu , Xiangyu Wu , Bo An This is my paper

Pith reviewed 2026-05-21 11:53 UTC · model grok-4.3

classification 💻 cs.RO cs.AI
keywords multi-robot path planningspatial attentionManhattan distancegraph attentionzero-shot generalizationdecentralized coordinationMAPPO
0
0 comments X

The pith

Embedding Manhattan distances into attention lets robots prioritize nearby teammates and reach 75 percent success when scaling from 8 to 128 agents.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper aims to show that standard attention in multi-robot communication wastes capacity on distant agents, so explicitly adding pairwise Manhattan distances to the weight calculation produces focused messages that improve coordination in crowded spaces. A reader would care because decentralized path planning must scale to large teams without central control or retraining if robots are to operate reliably in warehouses or disaster zones. The method introduces Relation enhanced Multi Head Attention that embeds distances, applies a distance mask, and fuses messages through a GRU before feeding into MAPPO training. Experiments demonstrate that this yields roughly 75 percent success on 40 by 40 grids with 30 percent obstacles when tested on 128 robots after training on only 8. Ablations isolate the distance encoding as the main source of the more than 25 point gain over baselines.

Core claim

RMHA embeds pairwise Manhattan distances directly into the attention weight computation so each robot can assign higher weight to messages from spatially close neighbors. Paired with a distance-constrained attention mask and GRU-gated fusion, the module plugs into MAPPO for end-to-end training. On 40 by 40 grids with 30 percent obstacle density, the resulting policy generalizes zero-shot from 8 training robots to 128 test robots at approximately 75 percent success rate, exceeding the strongest baseline by more than 25 percentage points. Ablation experiments identify the distance-relation encoding as the decisive factor behind gains in high-density regimes.

What carries the argument

Relation enhanced Multi Head Attention (RMHA), which augments standard multi-head attention by injecting pairwise Manhattan distances into the weight computation to let robots dynamically favor messages from nearby agents.

If this is right

  • Focused attention on nearby robots improves coordination precisely where obstacles force frequent local replanning.
  • The same architecture trains stably inside MAPPO without additional regularization.
  • Zero-shot scaling from 8 to 128 robots occurs without retraining or architecture changes.
  • Ablation results show that distance encoding, rather than extra heads or fusion alone, accounts for most of the observed lift.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same distance-aware attention pattern could be reused in other multi-agent settings that require selective long-range communication, such as formation control or distributed sensing.
  • Because distant messages are masked, total communication volume may drop in large swarms, offering a route to lower bandwidth use.
  • Physical-robot trials with noisy range sensors would test whether the learned prioritization survives the gap from simulation to hardware.

Load-bearing premise

That explicitly embedding pairwise Manhattan distances into the attention weight computation, combined with a distance-constrained mask and GRU fusion, will enable robots to prioritize spatially relevant neighbors sufficiently to produce the claimed performance gains in congested high-density environments.

What would settle it

Re-training and testing the identical MAPPO setup on the 128-robot 30-percent-obstacle cases but with the Manhattan-distance embedding removed from RMHA; if success rate then falls below 50 percent, the centrality of spatial embedding is falsified.

Figures

Figures reproduced from arXiv: 2603.02845 by Bo An, Sayang Mu, Xiangyu Wu.

Figure 1
Figure 1. Figure 1: Illustration of MRPP constraints: vertex collisions, edge collisions, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: MARL modeling paradigms for MRPP: MG, POMG, and Dec [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Architecture of the RMHA communication module. (1) Multi-modal observation encoder. A backbone of seven convolutional layers and two max-pooling layers ex￾tracts spatial features, followed by three fully connected layers for dimensionality reduction. For robot i at time t, the local observation o t i ∈ R 4×m×m is processed by: oˆ t i = Conv2 [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Network architecture of RMHA integrated with MAPPO. Algorithm 1 RMHA Training Procedure Input: Robot count N, episodes K, steps T, FOV F, layers L, discount γ, PPO clip ϵ 1: Initialize encoder ϕ, Transformer ψ, output heads θ, buffer B 2: for episode k = 0, . . . , K−1 do 3: for step t = 0, . . . , T −1 do 4: Encode observations via ϕ (Eqs. 8–10) 5: Compute messages and distances for all robots 6: for l = … view at source ↗
Figure 5
Figure 5. Figure 5: A 10×10 grid world with 30% obstacle density and 8 robots. Colored squares indicate start positions; colored circles indicate goals; gray circles represent obstacles; the red dashed box shows the 3×3 FOV. complexity increases to 15% and 30%, RMHA’s advantage becomes progressively larger. At 30%, RMHA maintains a success rate of approximately 75%, while SCRIMP falls below 50%, ODrM* drops to roughly 20% (wi… view at source ↗
Figure 8
Figure 8. Figure 8: Communication ablation: success rate comparison. [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Success rate comparison across algorithms on Warehouse and [PITH_FULL_IMAGE:figures/full_fig_p007_9.png] view at source ↗
Figure 12
Figure 12. Figure 12: Deadlock resolution in a narrow corridor. At [PITH_FULL_IMAGE:figures/full_fig_p008_12.png] view at source ↗
Figure 11
Figure 11. Figure 11: Global trajectory comparison: RMHA (left) vs. MAPPO (right). [2] J. Li, A. Tinka, S. Kiesel, et al., “Lifelong multi-agent path finding in large-scale warehouses,” in Proc. AAAI Conf. Artif. Intell., vol. 35, no. 13, 2021, pp. 11272–11281. [3] Y. Zhang, M. C. Fontaine, V. Bhatt, et al., “Multi-robot coordina￾tion and layout design for automated warehousing,” arXiv preprint arXiv:2305.06436, 2023. [4] K. D… view at source ↗
read the original abstract

Efficient communication is critical for decentralized Multi-Robot Path Planning (MRPP), yet existing learned communication methods treat all neighboring robots equally regardless of their spatial proximity, leading to diluted attention in congested regions where coordination matters most. We propose Relation enhanced Multi Head Attention (RMHA), a communication mechanism that explicitly embeds pairwise Manhattan distances into the attention weight computation, enabling each robot to dynamically prioritize messages from spatially relevant neighbors. Combined with a distance-constrained attention mask and GRU gated message fusion, RMHA integrates seamlessly with MAPPO for stable end-to-end training. In zero-shot generalization from 8 training robots to 128 test robots on 40x40 grids, RMHA achieves approximately 75 percent success rate at 30 percent obstacle density outperforming the best baseline by over 25 percentage points. Ablation studies confirm that distance-relation encoding is the key contributor to success rate improvement in high-density environments. Index Terms-Multi-robot path planning, graph attention mechanism, multi-head attention, communication optimization, cooperative decision-making

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes SPARC, a framework for decentralized multi-robot path planning that introduces Relation-enhanced Multi-Head Attention (RMHA). RMHA embeds pairwise Manhattan distances into attention weight computation, applies a distance-constrained attention mask, and uses GRU-gated message fusion, all integrated with MAPPO for end-to-end training. The central claim is strong zero-shot generalization: from 8 training robots to 128 test robots on 40x40 grids, RMHA reaches ~75% success rate at 30% obstacle density, outperforming the best baseline by >25 percentage points, with ablations attributing gains to the distance-relation encoding.

Significance. If the experimental claims hold under rigorous validation, the work would meaningfully advance learned communication for MRPP by addressing attention dilution in congested settings through explicit spatial relations. The zero-shot scaling result from 8 to 128 agents would be a notable contribution if supported by reproducible code, detailed baselines, and statistical analysis; the parameter-free nature of the Manhattan embedding is a potential strength worth highlighting.

major comments (2)
  1. [Abstract] Abstract: The headline performance numbers (~75% success rate at 30% density with 128 robots, >25pp improvement) are stated without any reference to experimental details, number of trials, error bars, statistical tests, baseline implementations, or specific tables/figures. This is load-bearing for the central generalization claim and prevents verification of whether the data support the result.
  2. [Abstract] Abstract (and method description): The distance-constrained attention mask is presented as key to prioritizing spatially relevant neighbors, yet no parameterization (e.g., radius value or adaptation rule), sensitivity analysis, or scaling behavior with agent count is supplied. A fixed radius tuned only on 8-robot training would risk either admitting too many neighbors (dilution) or too few (broken coordination) at 128-robot test density, directly undermining the zero-shot claim.
minor comments (1)
  1. [Abstract] The abstract mentions ablation studies confirming the importance of distance-relation encoding but provides no details on what was ablated or the quantitative impact; this should be expanded with specific results in the main text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. The comments highlight important aspects of clarity and rigor in presenting our experimental claims and methodological details. We address each major comment point-by-point below, indicating where revisions have been made to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The headline performance numbers (~75% success rate at 30% density with 128 robots, >25pp improvement) are stated without any reference to experimental details, number of trials, error bars, statistical tests, baseline implementations, or specific tables/figures. This is load-bearing for the central generalization claim and prevents verification of whether the data support the result.

    Authors: We agree that the abstract would benefit from explicit pointers to the supporting experimental evidence. In the revised manuscript, we have updated the abstract to reference Table II (which reports success rates with standard deviations over 100 independent trials), Figure 4 (showing statistical comparisons and p-values against baselines), and Section IV-B for the MAPPO training and zero-shot evaluation protocol on 40x40 grids at 30% obstacle density. These additions allow readers to directly verify the ~75% success rate and >25pp improvement claims. revision: yes

  2. Referee: [Abstract] Abstract (and method description): The distance-constrained attention mask is presented as key to prioritizing spatially relevant neighbors, yet no parameterization (e.g., radius value or adaptation rule), sensitivity analysis, or scaling behavior with agent count is supplied. A fixed radius tuned only on 8-robot training would risk either admitting too many neighbors (dilution) or too few (broken coordination) at 128-robot test density, directly undermining the zero-shot claim.

    Authors: We acknowledge that the original submission did not provide explicit parameterization details or sensitivity analysis for the distance-constrained mask. We have revised the method description in Section III-C to specify that the mask uses a fixed radius selected during 8-robot training to maintain local connectivity. To directly address the zero-shot scaling concern, we have added a new ablation subsection (Section V-D) with sensitivity results across multiple radius values and agent densities, along with analysis of attention distribution to show mitigation of dilution at 128 robots. The complementary pairwise Manhattan distance embedding in RMHA provides additional spatial weighting that supports generalization beyond the hard mask alone. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained empirical proposal

full rationale

The paper introduces RMHA as an architectural proposal that embeds pairwise Manhattan distances into attention weights, adds a distance-constrained mask, and fuses via GRU before training end-to-end with MAPPO. Performance numbers (75% success in zero-shot 8-to-128 robot transfer) are reported as measured outcomes of that training on 40x40 grids, not as quantities that are algebraically or statistically forced by the inputs. No equations or claims in the abstract reduce a prediction to a fitted parameter or to a self-citation chain; the distance embedding is presented as an explicit design choice whose value is then validated by ablation, not presupposed. The derivation chain therefore remains independent of its own outputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Based solely on the abstract, no explicit free parameters, axioms, or invented entities are described; the method extends standard attention and MAPPO without introducing new postulated entities.

pith-pipeline@v0.9.0 · 5705 in / 1117 out tokens · 55767 ms · 2026-05-21T11:53:36.666819+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.