acceptodds
Under review as a conference paper at ICLR 2027

Obstacle-Aware Multi-UAV Formation Coordination via a Hierarchical Guiding Vector Field with Adaptive Virtual Coordinates

Abstract

This paper presents a hierarchical guiding vector field (HGVF) with adaptive virtual coordinates for multi-UAV formation coordination in a two-dimensional workspace containing known static circular obstacles. The outer layer modifies the original closed path in its normal direction to construct an obstacle-aware coarse manifold. Based on the estimated obstacle risk, a shared formation variable continuously interpolates the tangential virtual-coordinate offsets and physical normal offsets between circular and line formation parameters. A distributed virtual-coordinate law coordinates the shifted virtual coordinates using neighboring scalar virtual-coordinate information. The inner layer guides each UAV along its individual reference and combines the resulting tracking field with local obstacle and inter-UAV repulsive corrections to generate bounded velocity commands. Numerical comparisons and CoppeliaSim results show positive obstacle clearances, temporary tracking and coordination deviations during formation reconfiguration, and transitions between the circular and line configurations in the considered scenario.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Multi-UAV formation coordination, hierarchical guiding vector field, adaptive virtual coordinates, formation reconfiguration, obstacle avoidance

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_GQTxOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies obstacle-aware formation coordination for multiple UAVs. The authors propose a hierarchical guiding vector field that combines obstacle-aware path deformation, adaptive formation reconfiguration, and virtual-coordinate coordination. The formation changes between circular and line configurations according to the estimated obstacle risk, while local repulsive terms are used for obstacle and inter-UAV avoidance. The method is evaluated through numerical simulations and CoppeliaSim experiments, including comparisons with two potential-field-based approaches.

Critical strengths

The overall framework is reasonable and addresses a practical problem in multi-UAV coordination. Combining path deformation with adaptive formation reconfiguration is useful for handling obstacles along a prescribed path. The numerical results also show improvements in obstacle clearance and reference tracking over the two compared baselines.

Critical weaknesses
  1. My main concern is the limited novelty. Guiding vector fields, virtual-coordinate coordination, and potential-field-based obstacle avoidance have been extensively studied. The proposed method mainly combines these existing techniques with a risk-dependent formation adjustment mechanism. I would also suggest moving the qualitative comparison in Table 1 from the Introduction to the Related Work section. More importantly, the comparison itself seems to suggest an incremental extension of existing methods rather than a substantial methodological advance. Although the overall framework is reasonable, I do not see a sufficiently new control principle or learning-based contribution. The distributed property is also limited, since the common formation variable still depends on information from all UAVs. Overall, the contribution appears rather limited, particularly for ICLR.

  2. The theoretical and experimental support is also insufficient. The paper does not establish convergence or collision-avoidance guarantees for the complete closed-loop system, especially during formation reconfiguration and under velocity saturation. The experiments are largely restricted to six UAVs moving in a two-dimensional environment with known static circular obstacles. There is no direct comparison with closely related guiding-vector-field methods, and no ablation study to distinguish the contributions of the main components. I also noticed inconsistent results between Tables 2 and 3 under seemingly identical parameter settings. These issues make it difficult to assess the reliability and generality of the reported improvements.

Questions for authors
  1. Could the authors clarify the main technical novelty compared with existing guiding-vector-field and virtual-coordinate coordination methods, particularly Yao et al. (2023)? Beyond combining established control techniques, what is the main generalizable contribution?

  2. Can the authors provide any convergence or collision-avoidance guarantees for the proposed closed-loop system, especially when the formation offsets vary with time and the velocity commands are saturated? If such guarantees are not available, what are the limitations of the method?

  3. Could the authors explain the differences between Tables 2 and 3 under the nominal parameter settings? It would also be useful to clarify why the experiments do not include direct comparisons with closely related GVF methods or ablations of the main components.

3 comments
AuthorsOpenReview ↗

Technical contribution and reuse

W1 Technical contribution and prior work

Review:

W1 “The proposed method mainly combines these existing techniques with a risk-dependent formation adjustment mechanism.”

Response: We use GVF to coordinate agents along a common route deformed from the known map. At each step, the same filtered map risk changes both the relative progress targets and the normal shift of each reference. The route itself stays fixed. The two offsets are aia_i and hih_i.

h_i=(1-\lambda)h_i^c+\lambda h_i^l,\quad p_i^d=f_o(w_i)+h_i n_o(w_i),\quad \Delta_{ij}=a_i-a_j.$$ Risk is evaluated at $\operatorname{mean}_i(w_i-a_i)$, not each robot's measured clearance. With $z_i=w_i-a_i$, the neighbor error is $z_i-z_j$. The update includes $(a_{i,k+1}-a_{i,k})/\Delta t$, cancelling the direct offset increment in the next offset-corrected coordinate when virtual clipping is inactive. Tracking uses the actual clipped rate $\nu_i$ in $\nu_i f_i'(w_i,t)-k_pe_i$. Repulsion and physical limiting modify the applied velocity. The GVF framework of [Yao et al. (2023)](https://doi.org/10.1109/TRO.2022.3224257) coordinates general parametric paths and surfaces. The main convergence analysis assumes fixed desired parameter differences and explicit graph and regularity conditions, although the general formulation permits time-varying differences. Sec. V provides a safety extension using quadratic programming (QP) and pairwise point distances, and Sec. VI analyzes heading alignment with saturated yaw input. Our addition is a rule that computes shared filtered risk from the map, adjusts both offsets together and includes the along-route offset increment in the coordination update. Yao et al. leave this rule unspecified. Their general path and surface representation can describe our references. Analyzing the time-varying closed loop here requires additional work. [Liu et al. (2026)](https://doi.org/10.1109/TAES.2026.3659453) combine formation surfaces defined along a route, route replanning and route-dependent scaling/rotation, and study tracking-coordination interaction. Our distinction is the specific shared risk rule, not greater expressiveness. Hu et al. study [spontaneous ordering](https://arxiv.org/html/2311.00976v1) and [flexible ordering](https://arxiv.org/html/2401.14013v1), and [Lu et al.](https://arxiv.org/html/2505.05840v1) study generalized formations. Our comparison with [Liu et al. (2025)](https://doi.org/10.1016/j.ifacol.2025.11.164) on obstacle-aware path following draws on their publisher abstract. **Connection to learning-based coordination.** In this known-map task, the risk thresholds and templates determine when the team starts changing shape and what references each member follows. This requires no policy training or fitted trajectory model. A learned controller could use these trajectories as references. That integration remains to be tested. **Q1 What can be reused and under what conditions** *Review:* Q1 “Beyond combining established control techniques, what is the main generalizable contribution?” **Response:** The offset rule can also use other routes or replace the circle and line templates. Desired differences are still computed by subtracting the two agents’ offsets. New templates still require checks on regularity, separation, clearance and executable rates, as well as closed-loop analysis. The original tracker omits explicit $\dot h_i n_o$ compensation, and nominal references during reconfiguration can coincide. Neighbor coordination exchanges scalar virtual coordinates, while shared risk needs aggregate progress and repulsion needs relative positions. The architecture is partially distributed, and network robustness remains untested. We will move Table 1 from the Introduction to Related Work to explain these mechanism and scope distinctions.
AuthorsOpenReview ↗

Safety and convergence

W2 Safety scope

Review:

W2 “The paper does not establish convergence or collision-avoidance guarantees for the complete closed-loop system”

Response: In Sec. 3.3, the applied velocity is the sum of tracking and both repulsive corrections after speed limiting. The reported clearances concern the tested planar point trajectories. A guarantee for the complete closed loop during reconfiguration requires a separate analysis of reference motion and both limits.

Q2 Sampled error dynamics and convergence limits

Review:

Q2 “especially when the formation offsets vary with time and the velocity commands are saturated?”

Response: Let qi,k=pid(wi,k,tk)q_{i,k}=p_i^d(w_{i,k},t_k), ei,k=pi,k−qi,ke_{i,k}=p_{i,k}-q_{i,k} and τ=Δt\tau=\Delta t. The raw velocity is ui,k=νi,kfi′(wi,k,tk)−kpei,k+ri,ku_{i,k}=\nu_{i,k}f_i'(w_{i,k},t_k)-k_p e_{i,k}+r_{i,k}, where νi,k\nu_{i,k} is the clipped virtual rate, fi′f_i' is the centered difference at fixed hih_i, and ri,kr_{i,k} is the sum of the two repulsive velocities. Let vi,kv_{i,k} be the applied velocity after the speed limit. Subtracting the next reference from the Euler position update gives

ei,k+1=(1−τkp)ei,k+τΦi,k.e_{i,k+1}=(1-\tau k_p)e_{i,k}+\tau\Phi_{i,k}.

Φi,k=νi,kfi′(wi,k,tk)+ri,k+(vi,k−ui,k)−qi,k+1−qi,kτ.\Phi_{i,k}=\nu_{i,k} f_i'(w_{i,k},t_k)+r_{i,k}+(v_{i,k}-u_{i,k})-\frac{q_{i,k+1}-q_{i,k}}{\tau}.

The reference increment includes normal-offset motion, which the tracking feedforward omits. Virtual-rate clipping also affects the shifted-coordinate update described in W1. Even when the scalar factor is contractive, the residual depends on the coupled closed loop. Zero-error convergence requires additional analysis.

These observations specify the scope of our conclusions. The distance, tracking and formation results describe behavior under the specified maps, parameters and initial conditions. They are not a general stability or safety theorem. We will make that distinction explicit in the method description and conclusion. The companion comment on Q3 and W2 gives the component and task evidence.

One possible certification direction is to check separation and obstacle clearance for all reference configurations during reconfiguration and combine a feasible safe initial set with constraints compatible with sampling and the applied velocity limits. Finite-body dynamics and uncertainty would require further analysis. These directions require a separate study and are not guarantees of the present controller.

AuthorsOpenReview ↗

Tables 2 and 3 and component tests

Q3 Tables 2 and 3 and the GVF comparison

Review:

Q3 “Could the authors explain the differences between Tables 2 and 3 under the nominal parameter settings?”

“It would also be useful to clarify why the experiments do not include direct comparisons with closely related GVF methods or ablations of the main components.”

Response: Tables 2 and 3. The discrepancy comes from three parameter settings omitted from the original manuscript. Table 2 used (kcol,Rcol,dline)=(1.00,1.00,m,0.50,m)(k_{\rm col},R_{\rm col},d_{\rm line})=(1.00,1.00\\,m,0.50\\,m). Table 3 used (0.45,0.72,m,0.58,m)(0.45,0.72\\,m,0.58\\,m). We checked all eight parameter combinations, including both original configurations, and reproduced their respective results from the same initial states. We also ran the original sensitivity-analysis code separately. Minimum sampled point distances were 0.27870147/0.06530318 m, and tracking RMSEs were 0.09497343/0.06546116 m. This explains the combined effect of the three parameter changes in that scene. For the proposed method and VAPF, the original coordination metrics satisfy RMSEc=(10.5m/rad)RMSEw\mathrm{RMSE}_c=(10.5\\ \mathrm{m/rad})\mathrm{RMSE}_w within a single run. This conversion cannot align different parameter configurations, and changing the offsets changes the reference task.

Comparison with Yao et al. Adding our routes, repulsion and limits to their GVF coordination law changes the method being tested. The experiments below measure the effects of our own components. A numerical comparison with Yao et al.’s complete method under matched conditions remains open. The theoretical comparison is in W1 and Q1.

W2 Component evidence and evaluation scope

Review:

W2 “There is no direct comparison with closely related guiding-vector-field methods, and no ablation study to distinguish the contributions of the main components.”

Response: In the earlier experiments with four environments and ten paired initial states per environment, fixed line has greater mean obstacle clearance and lower tracking error relative to its own reference than our complete controller (Full). Different reference tasks prevent using these errors as a ranking on an identical prescribed reference.

Reconfiguration task. All six points must pass the designated central gate without obstacle penetration, then keep the centered, labeled circular-formation RMS error within 0.09 m at the sampled instants over 1 s after all exit. No point-model trajectory may enter an obstacle during this interval. At 0.8/0.4 m, Full meets both conditions in 20/20 runs for each width. With the original tracking law, the fixed-circle condition has obstacle entry in 20/20 runs. Fixed line passes in 20/20 runs but recovers the required circular formation in 0/20 runs. Each condition and width uses two gate locations and ten paired starts. This supports reference reconfiguration for the stated point-model task, not a limitation of Yao's general framework or random-map generalization.

Component effects. Removing path deformation gives 40/40 runs entering an obstacle versus Full's 0/40. Removing inter-agent correction reduces mean minimum point distance from 0.05333 to 0.00357 m. Same-map risk lookup nearly reproduces Full, and a preset time schedule improves some geometric metrics. The tests therefore do not show an independent benefit from online static-map risk recomputation.

Obstacle-repulsion cost. In a separate experiment group with conflicting geometric requirements, repulsion on/off gives 0/40 runs entering an obstacle in both conditions. Mean minimum clearance is 0.53135/0.13904 m, but the number of events with line-template error at most 0.09 m at sampled instants over 1 s is 0/160 versus 150/160. These 160 events are repeated observations within 40 runs. They use different scenes and observation intervals from the older 130/160 line-then-circle result and cannot replace or be pooled with it.

Obstacle repulsion increases clearance but can impede shape transitions. These results support specific component and task conclusions for planar point agents in known static maps. They do not certify finite airframes or establish general controller superiority.

Reviewer_3kyJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a hierarchical guiding vector field (HGVF) framework for multi-UAV formation coordination in environments with known static obstacles. The method combines obstacle-aware path deformation, risk-driven formation reconfiguration through adaptive virtual coordinates, and local repulsive corrections for obstacle and inter-UAV avoidance.

The approach is evaluated through numerical simulations and CoppeliaSim experiments, including comparisons with two artificial potential field-based baselines and a parameter sensitivity analysis.

Critical strengths
  1. Clear problem formulation and structured methodology.The paper addresses a relevant multi-UAV coordination problem and organizes the proposed approach into interpretable components for path deformation, formation reconfiguration, and trajectory tracking.

  2. Preliminary empirical evidence.The simulation results demonstrate successful formation transitions and positive obstacle clearances in the evaluated scenario. The baseline comparisons and parameter sensitivity analysis provide some evidence of the method's effectiveness under the considered conditions.

Critical weaknesses
  1. Insufficient theoretical analysis of the proposed control framework. The paper clearly distinguishes its contributions from Yao et al. (2023), particularly through obstacle-aware path deformation and risk-driven, time-varying formation offsets. However, the theoretical properties of the resulting framework remain insufficiently investigated. In particular, the paper does not analyze closed-loop stability, tracking convergence, or collision avoidance during formation transitions. Given the introduction of time-varying references, local repulsive corrections, and velocity saturation, further theoretical justification or systematic analysis of the controller's behavior would strengthen the technical contribution.

  2. Limited experimental validation of the proposed mechanisms. The evaluation is primarily restricted to a single scenario involving six UAVs and four static circular obstacles. Although the results demonstrate the feasibility of the proposed approach, the experiments do not sufficiently establish the individual contributions of its key components. In particular, ablations of obstacle-aware path deformation, adaptive formation reconfiguration, and local repulsive corrections are missing. The quantitative comparisons are also limited to two artificial potential field-based methods, without a direct comparison against the closely related GVF coordination framework. Experiments with different obstacle configurations, UAV numbers, and initial conditions would provide stronger evidence of robustness and generalizability.

  3. Unclear experimental consistency and practical implementation. Tables 2 and 3 report noticeably different results under apparently identical default parameters. For example, the minimum inter-UAV distance differs substantially (0.27870 versus 0.065303), while the tracking RMSE also changes without an explanation of the experimental differences. Furthermore, although the authors clarify that the virtual-coordinate coordination term uses neighboring information, the formation-center estimation in Eq. (11) requires information from all UAVs, and the adaptive formation variable is shared globally. The practical communication requirements and scalability of this partially distributed architecture are not sufficiently evaluated.

Questions for authors
  1. Could the authors clarify the experimental differences between Tables 2 and 3? Given the stated default parameters, why do the minimum inter-UAV distances and tracking errors differ substantially?

  2. Could the authors provide ablation results isolating the effects of obstacle-aware path deformation and adaptive virtual-coordinate reconfiguration? Such results would help establish which components are responsible for the reported improvements.

  3. Given the absence of formal collision-avoidance guarantees, how should the reported minimum inter-UAV distances be interpreted in terms of physical UAV dimensions and collision safety?

4 comments
AuthorsOpenReview ↗

Closed-loop analysis and UAV dimensions

W1 Closed-loop analysis

Review:

W1 “the paper does not analyze closed-loop stability, tracking convergence, or collision avoidance during formation transitions.”

Response: The controller adds obstacle and inter-agent repulsion to the tracking command and limits the sum. The moving references and clipped virtual rates also affect the tracking error during reconfiguration. The observed point-model clearances therefore describe the tested runs. Proving collision avoidance and stability for the full controller requires a separate analysis.

We will state this scope in the method and conclusion. A possible analysis would check separation along the changing references and couple a feasible safe initial set with constraints compatible with the updates and speed limits.

Q3 Physical dimensions and collision interpretation

Review:

Q3 “how should the reported minimum inter-UAV distances be interpreted in terms of physical UAV dimensions and collision safety?”

Response: Original minima are point distances without airframe dimensions. The supplied static model is approximately 0.410×0.410×0.0300.410\times0.410\times0.030 m, with an enclosing sphere of radius 0.291 m centered at the root object's origin, checked through the simulator's geometry interface. If planar positions represent these roots at a common altitude, separation above 0.582 m suffices to separate the static envelopes. The reported minima 0.27870/0.065303 m do not satisfy this sufficient condition, so they do not certify body separation. Overlapping enclosing spheres alone do not prove body contact. Obstacle clearance already subtracts obstacle radius and remains about 0.69–0.70 m.

We therefore evaluated additional configuration B with the same control-law structure and speed cap, changing repulsion, tracking/transition parameters, circle radius/orientation, line spacing and deformation amplitude. B also uses a different feasible initial state, because the original initial minimum of 0.4673 m precludes envelope separation at t=0t=0.

We checked all 15 pairs throughout each held-velocity interval. The 0.06516 m value below is the interval minimum for Table 3. The original 0.065303 m value uses the sampled instants. Tracking RMSE uses each configuration’s own reference over 74 s.

Response Table 5. Minimum distances over held-velocity intervals and own-reference tracking RMSE.

ConfigurationMinimum pair distance (m)Obstacle-boundary clearance (m)Tracking RMSE (m)
Original Table 30.065160.692440.06546
Additional B0.676760.679860.08358

B has approximately 0.09476 m separation margin between the enclosing spheres. Under the stated static-body and piecewise-linear-motion assumptions, the enclosed bodies do not intersect in its nominal trajectory or the 20 recorded runs from the specified bounded set of initial states. Each additional run lasts 148 s and retains pair distance at least 0.69472 m. All nine nearby kλ/Ak_\lambda/A combinations and two step-size halvings meet the numerical targets. This is local validation in one static scene, not validation of the original parameter configurations or a universal avoidance theorem.

Greater inter-agent separation has a tracking cost: RMSE increases 27.7%, peak individual error rises from 0.382 to 0.609 m, and the RMS deviation from the labeled line template, after removing overall translation, is 0.303 m over samples with λ>0.9\lambda>0.9. The double-precision C++ implementation used for these runs was checked against complete archived MATLAB trajectories, with position differences below 10−810^{-8} m. The new configuration has not been rerun directly in MATLAB. Rotor sweep, aircraft dynamics and real-flight uncertainty remain uncertified.

AuthorsOpenReview ↗

Ablations and fixed-formation comparisons

W2 Components, GVF comparison and evaluation scope

Review:

W2 “ablations of obstacle-aware path deformation, adaptive formation reconfiguration, and local repulsive corrections are missing.”

“Experiments with different obstacle configurations, UAV numbers, and initial conditions would provide stronger evidence of robustness and generalizability.”

Response: In a separate 40-run test, obstacle repulsion increased mean minimum clearance from 0.13904 to 0.53135 m. Both conditions had 0/40 runs entering obstacles, and line-template attainment within 0.09 m for 1 s fell from 150/160 to 0/160 repeated events. Removing inter-agent correction in the component tests reduced mean minimum point separation from 0.05333 to 0.00357 m.

We tested four environments with ten paired initial states each. The gate study uses two locations and four widths. The initial-state draws are reused across the eight geometries. Both studies use six agents, 100 s and 0.02 s steps. Earlier 8/12-agent tests cover two layouts and three seeds each. The tests use first-order kinematics and known static maps.

Yao et al. (2023) analyze fixed desired differences under graph and regularity assumptions and describe point-distance barrier QPs and saturated heading control. Our controller uses a map-deformed route and risk-dependent offsets. A matched numerical comparison with Yao's complete method remains open.

Q2 What path deformation and reconfiguration contribute

Review:

Q2 “Could the authors provide ablation results isolating the effects of obstacle-aware path deformation and adaptive virtual-coordinate reconfiguration?”

Response: In the earlier 40 paired cases, point-model trajectories enter obstacles in 40/40 runs without path deformation, versus 0/40 for our complete controller (Full). Other modules remain active, and risk is recomputed on the undeformed route. The comparison measures the combined effects of changing the route geometry and recomputing risk. Fixed line nevertheless has better mean minimum clearance, point separation and RMSE relative to each method's own reference than Full.

The separate passage task tests reference reconfiguration. All six points must pass the central gate, then keep the centered, labeled circular-formation RMS error within 0.09 m at sampled instants over 1 s after all exit. No point-model trajectory may enter an obstacle during this interval. An exterior bypass fails the specified passage task but is not automatically a collision. For the tested fixed-shape circular template, rigid motion through the 0.8 m central gate is obstructed.

Response Table 6. Runs passing the gate, then runs completing both passage and recovery. Each cell uses two gate locations and ten paired starts.

Condition0.8 m gate0.4 m gate
Full20/20, 20/2020/20, 20/20
Fixed circle, original tracking0/20, 0/200/20, 0/20
Fixed line, original tracking20/20, 0/2020/20, 0/20

Full and fixed circle use the same tracking law and corrections, so this comparison supports reference reconfiguration for the specified task. Fixed line passes but misses the circular endpoint. In the wider 1.2/2.4 m cases, fixed circle also attains 20/20 with a shorter mean time to completion of both task requirements, while Full deforms unnecessarily for this task. The earlier tests also include cases without an obstacle-entry advantage for Full, near-equivalence of same-map lookup and lower geometric error or faster recovery of the circular formation with a preset time schedule. These results support performance on the specified task and its costs, not uniform tracking superiority or an independent benefit from recomputing static-map risk online. Sensing/coverage utility was not measured.

AuthorsOpenReview ↗

The reason of difference from tables 2 and 3 and communication

W3 Experimental consistency and communication

Review:

W3 “the formation-center estimation in Eq. (11) requires information from all UAVs, and the adaptive formation variable is shared globally.”

Response: The coordination step exchanges scalar coordinates with ring neighbors. Equation (11) also requires the progress of all agents, and every agent uses the same risk state. Repulsion needs relative positions. The table lists the required inputs per update.

The architecture is partially distributed. For NN agents, MM obstacles and coordination edges EE, each update uses:

Response Table 7. Information used per update.

ComponentRequired informationTeam computation
Virtual-coordinate coordinationRing-neighbor coordinates and shared formation offsetsO(N+∣E∣)=O(N)O(N+\lvert E\rvert)=O(N) on the ring
Center estimate, Eq. (11)All coordinates after subtracting the offsets wi−aiw_i-a_iO(N)O(N) aggregation
Shared formation stateAggregated center, static map and common λ\lambdaRisk O(M)O(M), member offsets O(N)O(N)
Inter-agent repulsionOther agents' relative positionsDirect pair checks O(N2)O(N^2)

The current simulation obtains this information from shared state arrays. Deployment would require aggregation, consistent sharing of the common state, and relative-position sensing or communication. The computation over all agent pairs grows quadratically, but does not determine packet counts, which depend on the protocol, broadcast and sensing arrangements. Bandwidth, delay, packet loss and distributed center estimation remain untested, so increasing the simulated agent count does not establish communication scalability.

Q1 Why Tables 2 and 3 differ

Review:

Q1 “Given the stated default parameters, why do the minimum inter-UAV distances and tracking errors differ substantially?”

Response: Three settings changed between Tables 2 and 3: the repulsion gain, repulsion range and nominal line spacing. The original manuscript omitted the differences. Using the two configurations reproduces both sets of values.

Response Table 8. Reproduced configurations and outcomes.

Setting/resultTable 2Table 3 default
Repulsion gain kcolk_{col}1.000.45
Repulsion range RcolR_{col} (m)1.000.72
Nominal line spacing dlined_{line} (m)0.500.58
Sampled minimum point separation (m)0.278701470.06530318
Position-reference tracking RMSE (m)0.094973430.06546116

Both runs use identical physical and virtual initial states from seed 1 and six agents, with the same path/map, kλ=1.05k_\lambda=1.05, A=1.75A=1.75, 74 s and 0.02 s steps. We checked all eight combinations of the three differing settings, including the two original parameter configurations, which reproduce their respective table values. We also ran the original sensitivity-analysis code once. Both metrics were independently recomputed from all nine saved trajectories. These checks support the combined effect of the three parameter changes for this scene and initial state. Because changing dlined_{line} changes the references, the RMSE values do not rank tracking of an identical reference. The prepared correction for Section 4.2 and the table captions lists both parameter configurations and retains their separate results.

I appreciate the authors' detailed rebuttal and additional experiments, which address several of my concerns. However, a direct comparison with the closely related GVF framework remains missing, making it difficult to assess the proposed method's incremental contribution. Therefore, I maintain my rating.

Reviewer_hgDtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

Obstacle-Aware Multi-UAV Formation Coordination via a Hierarchical Guiding Vector Field with Adaptive Virtual Coordinates This paper studies multi-UAV formation coordination along a closed path in a 2D workspace with static circular obstacles. It proposes a hierarchical guiding vector field (HGVF) with adaptive virtual coordinates: the outer layer deforms the base path to construct an obstacle-aware reference, a risk-driven variable λ(t) interpolates between circular and line formations, the middle layer coordinates virtual coordinates through neighbor-only terms, and the inner layer tracks individual references with local obstacle and inter-UAV repulsive corrections, producing bounded velocity commands. Experiments with 6 UAVs on a cycle graph over 74 s with 4 obstacles compare against VAPF and an improved APF, with a parameter sensitivity analysis, in numerical simulation and CoppeliaSim.

Critical strengths

None

Critical weaknesses

None

Questions for authors
  1. There is no safety analysis. This is a safety-critical multi-UAV problem, but the paper provides no formal guarantee of collision avoidance or closed-loop stability. The repulsive terms are explicitly “local corrections rather than formal safety guarantees” (Sec. 3.3); the coupled closed-loop system is never analyzed, and the interplay between the repulsive terms (28-29) and the tracking field is not discussed.
  2. The related-work survey does not engage with recent top-conference research relevant to ICLR. The survey cites exclusively control and robotics conferences, without a single reference from ICLR, NeurIPS, ICML, CVPR, or other top conferences, nor any discussion of how the proposed method relates to recent work published at these conferences. As a result, the work does not position its contribution against the latest advances in the ICLR research ecosystem, which weakens both the novelty claim and the venue fit.
  3. Insufficient baselines and no ablation. Only VAPF and an improved APF are compared, and no numerical comparison with the most relevant GVF-based work (Yao et al., 2023) is provided. Comparisons with fixed circular and line formations are missing.
  4. Scope fit for ICLR is not argued. The work is a deterministic, model-based feedback controller with fixed design gains and no learning component. It does not establish why the problem or results are of interest to the ICLR community, nor does it discuss connections to learning-based multi-agent coordination.
  5. The adaptive mechanism is under-specified: the mapping from obstacle risk to λ(t) is given only by formulas without parameter values, and b(w) is a nominal normal displacement that should be distinguished from the risk-driven offsets.
  6. Unclear experimental metrics: RMSEc and RMSEp are not explicitly defined in Sec. 4.2, and Table 3 introduces RMSEw without clarifying its relation to the other metrics.
  7. In Sec. 3.1, “The obstacle-aware obstacle-aware path is defined by” contains a duplicated word.
  8. CoppeliaSim results (Fig. 4) only show motion snapshots without quantifying agreement with the numerical simulation.
4 comments
AuthorsOpenReview ↗

Safety and closed-loop stability

Q1 Research focus and formal safety analysis

Review:

Q1 “the coupled closed-loop system is never analyzed, and the interplay between the repulsive terms (28-29) and the tracking field is not discussed.”

Response: This paper constructs formation references for a known obstacle map and coordinates how the agents follow those references. In Sec. 3.3, we add the obstacle and inter-agent repulsive velocities in Eqs. (28–29) to the tracking command, then limit the sum. During reconfiguration the reference also moves, so a tracking-only argument would leave full closed-loop stability open.

A formal collision-avoidance proof for the complete controller is a separate research task. The reported clearances describe the tested point-model runs. We will state that scope in Sec. 3.3 and the conclusion. A follow-up would first check how reference clearance and separation change throughout a transition, then impose safety constraints compatible with the sampling period and applied speed limits. Finite airframes and flight dynamics require corresponding models and validation.

AuthorsOpenReview ↗

Related work and ICLR relevance

Q2 Related work

Review:

Q2 “The related-work survey does not engage with recent top-conference research relevant to ICLR.”

Response: STAF (CoRL 2025) learns subteaming and adaptation within a formation type. DGPPO (ICLR 2025) learns a distributed policy and discrete graph barriers. MMD (ICLR 2025) and SMD (ICML 2025) use diffusion models to generate multi-robot trajectories. In our task, one team tracks explicit geometric references that change from a circle to a line.

Yao et al. (2023) provide the virtual-coordinate GVF framework. Hu et al. study ordering and surface navigation, and Lu et al. study generalized formations. The abstract of Liu et al. (2025) describes coordinated path following around obstacles. We use a fixed, map-deformed route and a common risk value to adjust the two offsets together. We compute risk from the map at the estimated reference center. Q3 explains how the coordination update differs from Yao's formulation.

Q4 What advantages does this method offer over learned approaches?

Review:

Q4 “It does not establish why the problem or results are of interest to the ICLR community, nor does it discuss connections to learning-based multi-agent coordination.”

Response: Given the static map, we compute the route and use obstacle risk to adjust the reference geometry. This has three practical consequences.

  1. The controller uses the map and current state directly. It needs no policy-training trajectories, training reward or fitted policy.
  2. The risk thresholds and formation templates specify when the formation starts changing and what geometry it approaches. As risk rises, the normal offsets shrink toward zero and the references approach a line. The resulting targets can be checked directly.
  3. For another known map, we can recompute the references with the same equations and check whether the templates and gains are feasible for that scene. There are no trained weights to refit.

STAF and our work both address how agents coordinate as obstacles require a change of formation. The generated trajectories could serve as references for a learned multi-agent controller. A matched experiment could also test how the approaches adjust formation on the same maps. These are follow-up studies. Our current evidence concerns first-order kinematics on known static maps.

AuthorsOpenReview ↗

Comparison with other method and ablations

Q3 Theoretical distinction from Yao et al.

Review:

Q3 “no numerical comparison with the most relevant GVF-based work (Yao et al., 2023) is provided. Comparisons with fixed circular and line formations are missing.”

Response: Yao et al. (2023) formulate GVF coordination on general parameterized paths and surfaces with possibly time-varying desired parameter differences. Their main theorems address fixed differences under graph connectivity and regularity assumptions.

We use the known map to deform a common route before each run. The route then remains fixed. A shared filtered risk λ(t)\lambda(t) adjusts along-route offsets ai(t)a_i(t) and normal offsets hi(t)h_i(t), yielding pid=fo(wi)+hi(t)no(wi)p_i^d=f_o(w_i)+h_i(t)n_o(w_i) and Δij(t)=ai(t)−aj(t)\Delta_{ij}(t)=a_i(t)-a_j(t). The virtual update adds the along-route offset increment. When the virtual rate is unclipped, subtracting the next offset cancels this increment.

Yao's general framework can represent our references. The paper leaves this particular map-dependent construction unspecified, and its fixed-difference theorems cannot establish convergence of our time-varying closed loop. We will clarify the distinction in Related Work and the method.

Experimental results for Q3

We tested fixed circle, fixed line and component removal within our controller. A matched numerical comparison with Yao's complete method remains open. Full includes path deformation, reconfiguration and local repulsion.

M1. Fixed formations. Six planar single-integrator points run for 100 s at 0.02 s steps. All points must pass the central gate without entering obstacles. Full task success also requires centered, labeled circular-formation RMS error at most 0.09 m at each sample over 1 s after all exit, with no obstacle entry during that interval. Each condition and width uses two gate locations and ten paired initial states, for 20 runs. Each set of initial states is shared across controllers and reused across designed geometries.

Response Table 1. Passages and complete tasks at the two narrow widths.

Reference condition0.8 m passage / complete task0.4 m passage / complete task
Full, original tracking law20/20, 20/2020/20, 20/20
Fixed circle, original tracking law0/20, 0/200/20, 0/20
Fixed line, original tracking law20/20, 0/2020/20, 0/20

The gate uses two opposing radius-6 m obstacles. Only central passage counts. For the tested fixed-shape circular template, rigid motion along the central branch is obstructed at 0.8 m. At wider 1.2/2.4 m gates, fixed circle completes the whole task in 20/20 runs and has a shorter mean completion time than Full.

M2. Component metrics in separate experiment groups.

Response Table 2. Each row compares Full with the stated modification in the corresponding experiment.

MetricFullModified condition
Runs entering an obstacle in the point model, path deformation removed0/4040/40
Mean minimum point separation, inter-agent correction removed (m)0.053330.00357
Runs entering an obstacle in the point model, obstacle repulsion off0/400/40
Mean minimum obstacle clearance, obstacle repulsion off (m)0.531350.13904
Events with line-template error at most 0.09 m at every sample over 1 s, obstacle repulsion off0/160150/160

The 160 events come from 40 runs. Removing path deformation also recomputes risk. Earlier fixed-line runs have better mean clearance, separation and RMSE against their own references. These references differ from Full's.

M3. Tracking-controller check. Both controllers complete the task in 20/20 runs per width. Mean times run from first entry to the end of the required circular-formation recovery interval.

Response Table 3. Mean time to finish the passage and recovery task.

Controller0.8 m (s)0.4 m (s)
Full17.1119.99
Separately constructed compensated extension14.7315.32

The compensated extension shows that a different tracking law also completes this task. Same-map risk lookup at actual virtual progress nearly reproduces Full within interpolation error. A preset schedule improves some geometric metrics and speeds up circle recovery. The tests cover planar points and leave the independent benefit of online risk recomputation open.

AuthorsOpenReview ↗

Risk parameters, metrics and simulator agreement

Q5 Risk mapping and the two offsets

Review:

Q5 “the mapping from obstacle risk to λ(t) is given only by formulas without parameter values”

“b(w) is a nominal normal displacement that should be distinguished from the risk-driven offsets.”

Response: We deform the route once. The two formation offsets change with risk during the run. The fixed route is fo(w)=fb(w)+b(w)nb(w)f_o(w)=f_b(w)+b(w)n_b(w) with b(w)=∑m(−sm)Amexp⁡[−12(wrap⁡(w−wm)/σm)2]b(w)=\sum_m(-s_m)A_m\exp[-\tfrac12(\operatorname{wrap}(w-w_m)/\sigma_m)^2]. The scene generator presets the anchors and obstacle sides sms_m relative to nbn_b. wrap⁡(δ)=mod⁡(δ+π,2π)−π\operatorname{wrap}(\delta)=\operatorname{mod}(\delta+\pi,2\pi)-\pi.

Risk adjusts ai=(1−λ)aic+λaila_i=(1-\lambda)a_i^c+\lambda a_i^l and hi=(1−λ)hic+λhilh_i=(1-\lambda)h_i^c+\lambda h_i^l. In pid=fo(wi)+hino(wi)p_i^d=f_o(w_i)+h_i n_o(w_i), bb moves the common route along nbn_b, hih_i shifts the reference along non_o and aia_i sets relative virtual-coordinate targets.

Let C(x)=min⁡(1,max⁡(0,x))C(x)=\min(1,\max(0,x)), w^c=N−1∑i(wi−ai)\hat w_c=N^{-1}\sum_i(w_i-a_i) and p^c=fo(w^c)\hat p_c=f_o(\hat w_c). This locates the estimated center on the route. With dmb=∣p^c−cm∣−(rm+db)d_m^b=\\|\hat p_c-c_m\\|-(r_m+d_b),

\rho=C\left(\max_m\max\left\\{C((R_\rho-d_m^b)/R_\rho),\\ \kappa_\rho e^{-\frac12(\operatorname{wrap}(\hat w_c-w_m)/\sigma_\rho)^2}\right\\}\right).

Nominal settings are Am=1.75A_m=1.75 m, σm=0.135\sigma_m=0.135 rad, db=0.35d_b=0.35 m, Rρ=2.50R_\rho=2.50 m, σρ=0.17\sigma_\rho=0.17 rad, κρ=0.80\kappa_\rho=0.80, (ρl,ρh)=(0.22,0.72)(\rho_l,\rho_h)=(0.22,0.72), kλ=1.05k_\lambda=1.05 s$^{-1}$, Δt=0.02\Delta t=0.02 s and λ0=0\lambda_0=0. The target λd\lambda_d is 0 at or below ρl\rho_l, 1 at or above ρh\rho_h, and 10s3−15s4+6s510s^3-15s^4+6s^5 between them, where s=(ρ−0.22)/0.50s=(\rho-0.22)/0.50. We update λk+1=C(0.979λk+0.021λd,k)\lambda_{k+1}=C(0.979\lambda_k+0.021\lambda_{d,k}).

Templates use N=6N=6, Rf=0.90R_f=0.90 m, sf=10.5s_f=10.5 m/rad and θb=π/6\theta_b=\pi/6. The main and sensitivity runs use dl=0.58d_l=0.58 m and the original comparison uses 0.50 m. The increment dl/sfd_l/s_f sets virtual spacing. Secs. 3.1–3.2 and 4.1 will list these settings.

Q6 Metric definitions and correction

Review:

Q6 “RMSEc and RMSEp are not explicitly defined in Sec. 4.2”

Response: Table 2 scales our and VAPF’s virtual errors to metres. Table 3 reports ours in radians. APF uses physical relative-reference error. For the 74 s including initialization and reconfiguration, let A[x]=(0.02/74)(x0/2+∑k=13699xk+x3700/2)\mathcal A[x]=(0.02/74)(x_0/2+\sum_{k=1}^{3699}x_k+x_{3700}/2). With each method's reference qiq_i, use ei=pi−qie_i=p_i-q_i, δij=(wi−ai)−(wj−aj)\delta_{ij}=(w_i-a_i)-(w_j-a_j) and the six undirected ring edges EE, counted once each.

\mathrm{RMSE}_w=\sqrt{\mathcal A[|E|^{-1}\sum_{(i,j)\in E}\delta_{ij}^2]}.$$ Figure 2(b) shows all 15 pairs. We use Euclidean distances and unwrapped coordinates. For our method and VAPF, $\mathrm{RMSE}_c=(10.5\\ \mathrm{m/rad})\mathrm{RMSE}_w$ within a run. The tables use different configurations. We withdraw the mixed-metric ranking and use $$\mathrm{RMSE}_{\mathrm{rel}}=\sqrt{\mathcal A[|E|^{-1}\sum_{(i,j)\in E}\\|(p_i-p_j)-(q_i-q_j)\\|^2]}.$$ Thirty saved trajectories (three methods, seeds 101–110, $k_{col}=0.45$, $R_c=0.72$ m, $d_{line}=0.58$ m) give the following results over the first 74 s. **Response Table 4.** Mean ± sample standard deviation. | Method | RMSEp (m) | RMSErel (m) | RMSEw (rad) | |---|---:|---:|---:| | Proposed | 0.065473 ± 0.000052 | 0.076069 ± 0.000142 | 0.012321 ± 0.000797 | | VAPF | 0.345639 ± 0.000010 | 0.404835 ± 0.000026 | 0.022428 ± 0.000444 | | APF | 0.149496 ± 0.000776 | 0.136629 ± 0.001667 | N/A | We will list the metric definitions in Sec. 4.2. **Q7 Duplicated wording** *Review:* Q7 “The obstacle-aware obstacle-aware path is defined by” contains a duplicated word. **Response:** The corrected wording for Sec. 3.1 is “The obstacle-aware path is defined by”. **Q8 Quantitative CoppeliaSim agreement** *Review:* Q8 “CoppeliaSim results (Fig. 4) only show motion snapshots without quantifying agreement with the numerical simulation.” **Response:** The paired numerical and CoppeliaSim 4.7.0 rev4 runs have position RMS differences of 1.60/1.49/1.53 μm for seeds 1/2/3, with maxima of 4.57/4.42/4.38 μm. Tracking RMSE averages 0.06543 m on both sides. The largest per-seed difference is $4.95\times10^{-9}$ m. Both branches use Table 3 settings, six agents, 3,700 steps over 74 s and the same MATLAB controller. CoppeliaSim integrates the first-order model and returns positions for feedback. After a client export error, we reconstructed controller histories from logged positions rounded to float32 and checked every logged velocity command. These runs quantify kinematic agreement and complement Fig. 4.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.