Paper deep dive
CollisionGAT: Controller-Agnostic One-Step Collision Screening for Multi-Agent Motion
Alan Debbas, Edwin Meriaux, Gregory Dudek
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 9/29/2026, 11:12:00 AM
Summary
The paper introduces CollisionGAT, a graph-attention network designed for controller-agnostic, one-step collision screening in multi-agent motion systems. It processes current and proposed states of agents and obstacles to output collision-risk scores, enabling controllers to accept, repair, or replan steps. The authors demonstrate its integration with a continuous path-following controller and GATeD, a discrete planner based on D* Lite, achieving high precision and recall in both continuous and grid-based environments.
Entities (8)
Relation Signals (6)
CollisionGAT → isa → Graph Attention Network
confidence 95% · We present CollisionGAT, a graph-attention network that reads the current and proposed states of moving agents
CollisionGAT → isusedby → GATeD
confidence 95% · We mount CollisionGAT on a continuous path-following controller and on GATeD
Alan Debbas → affiliatedwith → McGill University
confidence 90% · Alan Debbas Affiliation: McGill University Montreal, Canada
Edwin Meriaux → affiliatedwith → McGill University
confidence 90% · Edwin Meriaux Affiliation: McGill University Montreal, Canada
Gregory Dudek → affiliatedwith → McGill University
confidence 90% · Gregory Dudek Affiliation: McGill University Montreal, Canada
GATeD → isbasedon → D* Lite
confidence 90% · GATeD, an obstacle-blind D* Lite planner
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Before a team of robots moves, each proposed step must be checked for collisions with other robots and with obstacles. We present CollisionGAT, a graph-attention network that reads the current and proposed states of moving agents together with locally relevant stationary obstacles and returns one collision-risk score per moving agent. Any controller can use these scores to accept, repair, replan, or postpone a proposed step. We mount CollisionGAT on a continuous path-following controller and on GATeD, an obstacle-blind D* Lite planner that uses typed vetoes to update its planning graphs. Exact geometric checks supply the training labels and independently audit every executed step.
Tags
Links
- Source: https://arxiv.org/abs/2609.32783v1
- Canonical: https://arxiv.org/abs/2609.32783v1
Trouble viewing inline? Open PDF directly →
Full Text
29,237 characters extracted from source content.
Expand or collapse full text
CollisionGAT: Controller-Agnostic One-Step Collision Screening for Multi-Agent Motion Alan Debbas Affiliation: McGill University Montreal, Canada Alan.Debbas@mail.mcgill.ca Edwin Meriaux Affiliation: McGill University Montreal, Canada Edwin.Meriaux@mail.mcgill.ca Gregory Dudek Affiliation: McGill University Montreal, Canada Gregory.Dudek@mcgill.ca Abstract Before a team of robots moves, each proposed step must be checked for collisions with other robots and with obstacles. We present CollisionGAT, a graph-attention network that reads the current and proposed states of moving agents together with locally relevant stationary obstacles and returns one collision-risk score per moving agent. Any controller can use these scores to accept, repair, replan, or postpone a proposed step. We mount CollisionGAT on a continuous path-following controller and on GATeD, an obstacle-blind D∗ Lite planner that uses typed vetoes to update its planning graphs. Exact geometric checks supply the training labels and independently audit every executed step. Index Terms: multi-agent systems, collision detection, graph attention networks, motion planning, D* Lite I Introduction Before several robots move together, a controller must decide whether their proposed next steps can be executed without collision. The decision is relational: a step that is safe for one robot alone can land on another robot’s destination, enter an obstacle, or send two robots to the same grid cell. The number of neighbors that matter changes at each timestamp, so the checker must handle a variable number of agents and obstacles. Exact geometric predicates supply training labels and audit execution, but exhaustive screening evaluates up to (n2)+nm n2+nm agent–agent and agent–obstacle pairs. Before inference, each proposed motion segment is enclosed in an axis-aligned box expanded by the agent radius; only sender pairs whose boxes can overlap are retained. CollisionGAT batches these pairs without receiving relative-distance or collision-margin features. The interface is simple: each agent proposes one next state, CollisionGAT returns one risk score per proposal, and the controller decides whether to commit, repair, replan, or wait. Our contributions are (i) a complete input–output contract for a four-head pairwise-attention collision classifier, (i) a training progression from small collision primitives to full task-scale graphs, and (i) a continuous and a discrete controller built on the same interface with separately trained weights. I Background CollisionGAT draws on three lines of work: graph attention networks, learned safety checking for multi-agent motion, and incremental path planning. Graph attention networks learn how much each neighbor should matter [1], and dynamic pairwise attention lets each receiver rank the same neighbors differently [2]. CollisionGAT uses this relational machinery for a narrow task: classifying each agent’s proposed step as risky or safe. Learned models have been used to accelerate collision queries in motion planning [3]. Control-admissibility models score sampled actions with a graph network while a separate function drives progress toward the goal [4]. GCBF+ jointly learns a graph barrier function and a distributed controller [5]. CollisionGAT deliberately does less: the external controller supplies the action, the network only classifies it, and masked sender views attribute a grid veto to an agent or obstacle. Control synthesis and formal certification remain complementary to this checking role. Our grid controller builds on D∗ Lite, which incrementally repairs shortest paths when edge costs change [6]. Deterministic priority rules resolve simultaneous agent conflicts, while the frozen checker supplies the online graph updates. I Problem Formulation Definition 1 (World and active agents): Let W be a world containing a finite initial set of mobile agents 0A_0 and stationary obstacles O. Agent i has start i∈ s_i and goal i∈ g_i . At timestamp t, the active set t⊆0A_t _0 contains the agents that have not yet reached their goals. Active agent i occupies i p_i, and the controller proposes a one-step state i q_i. Arrival is absorbing: when agent i arrives, t←t∖iA_t _t \i\. Unsubscripted bold symbols such as =(i)i∈t p=( p_i)_i _t stack the indexed states. In continuous motion, =[0,L]2⊂ℝ2W=[0,L]^2 ^2 and i=i+ℓii∈ q_i= p_i+ _i d_i , where ℓi≥0 _i≥ 0 and ‖i‖2=1\| d_i\|_2=1 when ℓi>0 _i>0. Agent i occupies the closed ball B¯(z,ri):=x:‖x−z‖2≤ri B(z,r_i):=\x:\|x-z\|_2≤ r_i\ with ri>0r_i>0, and both the current and the proposed footprint lie inside W. Each obstacle o∈o is a pair (o,ao)( c_o,a_o) with ao>0a_o>0 and occupies the closed square Q(o,ao):=x:max(|x1−co,1|,|x2−co,2|)≤aoQ( c_o,a_o):=\x: (|x_1-c_o,1|,|x_2-c_o,2|)≤ a_o\. In discrete motion, =ΛN:=0,…,N−12W= _N:=\0,…,N-1\^2, the obstacle set is a subset of grid cells, starts are pairwise distinct. We define the starting position and end goal of agent i as i,i∈ΛN∖ s_i, g_i∈ _N . A legal proposal either stays in place or moves to one of the four neighboring cells: 4(c):=c′∈ΛN:‖c′−c‖1=1,i∈i∪4(i).N_4(c):=\c ∈ _N:\|c -c\|_1=1\, q_i∈\ p_i\ _4( p_i). (1) i p_ij p_ji q_ij q_j(a) endpoint overlap:detected by swept label×i p_ij p_ji q_ij q_j(b) crossing, safe endpoints:detected by swept labeliijji=j q_i= q_j(c) duplicate destination:detected by grid labeliijj×i=j,j=i q_i= p_j,\ q_j= p_i(d) head-on swap:detected by grid label Fig. 1: Agent–agent collision cases in continuous and grid motion. Dashed circles indicate current positions and solid discs or cells indicate proposed positions. The continuous swept label detects (a) endpoint overlap (subcase of a crossing overlap) and (b) crossing between safe endpoints. Under four-connected grid motion, (c) duplicate destinations and (d) head-on swaps are the only possible one-step agent collisions; every participant is labeled. The next two definitions give the exact collision checks. They supply all training labels and independently re-check every executed step; Fig. 1 illustrates the agent cases. Throughout, P=11\P\=1 when proposition P is true and 00 otherwise, and a maximum over an empty set is zero. Definition 2 (Grid labels): A step collides when it enters an obstacle cell, or when two agents claim the same cell or swap head-on (Fig. 1c,d): yiobs y_i obs =i∈, =1\ q_i \, (2) yiagt y_i agt =∃j∈t∖i: =1\! \∃ j _t \i\: . i=j∨(i=j∧j=i). 22.76219pt . q_i= q_j\ \ ( q_i= p_j q_j= p_i) \. (3) For the grid model, yiall=yiagt∨yiobsy_i all=y_i agt y_i obs. Definition 3 (Swept-motion labels): In continuous motion, agent i follows the straight center trajectory γi(λ)=(1−λ)i+λi,λ∈[0,1]. _i(λ)=(1-λ) p_i+λ q_i, λ∈[0,1]. Its swept agent-collision label is yisw,agt=maxj∈t∖iminλ∈[0,1]∥γi(λ)−γj(λ)∥2≤ri+rj.y_i^sw,agt= _j _t \i\1\! \ _λ∈[0,1]\| _i(λ)- _j(λ)\|_2≤ r_i+r_j \. (4) For a point x and a closed square Q, let dist(,Q)=min∈Q‖−‖2dist( x,Q)= _ z∈ Q\| x- z\|_2. The swept obstacle-collision label is yisw,obs=maxo∈minλ∈[0,1]dist(γi(λ),Q(o,ao))≤ri.y_i^sw,obs= _o 1\! \ _λ∈[0,1]dist ( _i(λ),Q( c_o,a_o) )≤ r_i \. (5) For the continuous model, yisw,all=yisw,agt∨yisw,obs.y_i^sw,all=y_i^sw,agt y_i^sw,obs. Any controllerunchangedCollisionGATone risk score per agentExecuteonly vetted actionsproposals qsafeveto: repair and re-proposenext timestamp Fig. 2: CollisionGAT attaches between any controller and execution: safe joint proposals pass through, and a veto sends the controller back to repair and re-propose. The controller itself is unchanged. IV CollisionGAT CollisionGAT sits in conjunction with an external controller. The controller sends in the current and proposed states of every node, each typed as an agent or an obstacle, and CollisionGAT returns one collision-risk score per active agent. This section answers three questions: what the network looks at (the interaction graph), what numbers describe each entity (the features), and how those numbers become a risk score (the classifier). This model can be seen in Fig. 2. IV-A Interaction graph and minimal features Definition 4 (Continuous and grid interaction graphs): Let χt(u,i) _t(u,i) be a binary edge filter. We set χt(u,i)=1 _t(u,i)=1 when sender u could collide with agent i during the proposed step, and χt(u,i)=0 _t(u,i)=0 otherwise. Let ℓmax _ be the largest continuous candidate step. For continuous motion, tcont _t cont =o∈:∃i∈t,dist(i,Q(o,ao))≤ℓmax+ri, = \o : subarrayc∃ i _t,\\ dist( p_i,Q( c_o,a_o))≤ _ +r_i subarray \, tcont _t cont =t⊔tcont, =A_t _t cont, ℰtcont _t cont =(u,i):i∈t,u∈tcont∖i,χt(u,i)=1. = \(u,i): subarrayci _t,\\ u _t cont \i\,\ _t(u,i)=1 subarray \. (6) For grid motion, tgrid _t grid =∩⋃i∈t(i∪4(i)), =O∩ _i _t (\ p_i\ _4( p_i) ), tgrid _t grid =t⊔tgrid, =A_t _t grid, ℰtgrid _t grid =(u,i):i∈t,u∈tgrid∖i. = \(u,i): subarrayci _t,\\ u _t grid \i\ subarray \. (7) Definition 5 (Raw states and sender masks): For continuous scenes, let t c_t be the centroid (average position) of the active-agent positions. The raw seven-dimensional agent and obstacle rows are i x_i =[i−t‖i−t‖ℓi‖ri‖0], =[ p_i- c_t\| q_i- c_t\| _i\|r_i\|0], (8) o x_o =[o−t‖o−t‖0‖ao‖1]. =[ c_o- c_t\| c_o- c_t\|0\|a_o\|1]. Each row contains the centered current and proposed positions, step length, entity size, and a binary obstacle indicator. Obstacles are stationary, so their positions are repeated and their step length is zero. Centering removes global translation. For a grid, let N=((N−1)/2,(N−1)/2) c_N=((N-1)/2,(N-1)/2) and eN(c)=(c−N)/50e_N(c)=(c- c_N)/50. The origin may lie between cells when N is even; the divisor is half the largest trained side and only stabilizes magnitude. The raw five-dimensional rows are i=[eN(i)‖eN(i)‖0],o=[eN(o)‖eN(o)‖1]. x_i=[e_N( p_i)\|e_N( q_i)\|0], x_o=[e_N(o)\|e_N(o)\|1]. (9) Neither model receives engineered relative displacement, distance, or collision-margin features. Stacking the rows gives Xtcont∈ℝ|tcont|×7X_t cont ^|V_t cont|× 7 or Xtgrid∈ℝ|tgrid|×5X_t grid ^|V_t grid|× 5; we write XtX_t when the setting is clear. IV-B Four-head pairwise-attention classifier Definition 6 (CollisionGAT map): We follow standard graph-attention networks terminology and theory [1, 2]. For a receiving agent i and a sending node u, the pair representation is ui=[u∥i]∈ℝ2d z_ui=[ x_u\| x_i] ^2d, where d∈5,7d∈\5,7\. For each attention head h∈1,…,4h∈\1,…,4\, the learned MLP ϕah:ℝ2d→ℝ _a^h:R^2d maps ui z_ui to a scalar attention logit, while ϕmh:ℝ2d→ℝ16 _m^h:R^2d ^16 maps it to the message sent from u to i. For a sender view μ∈all,agt,obsμ∈\all,agt,obs\, define: ℰtall=ℰt,ℰtagt=(u,i)∈ℰt:u∈t,ℰtobs=(u,i)∈ℰt:u∈t,iμ=u:(u,i)∈ℰtμ. aligned E_t all&=E_t,\\ E_t agt&=\(u,i) _t:u _t\,\\ E_t obs&=\(u,i) _t:u _t\, aligned _i^μ=\u:(u,i) _t^μ\. If the set is empty, set iμ= M_i^μ= 0; otherwise, each head takes the attention-weighted average of its messages: iμ=∥h=14∑u∈iμeϕah(ui)∑v∈iμeϕah(vi)ϕmh(ui)∈ℝ64. M_i^μ= \|_h=1^4 _u _i^μ e _a^h( z_ui) _v _i^μe _a^h( z_vi) _m^h( z_ui) ^64. (10) Different heads may emphasize different interactions. Let ϕx:ℝd→ℝ64 _x:R^d ^64 be the self encoder, ϕu:ℝ128→ℝ64 _u:R^128 ^64 the update map, and ∈ℝ64 w ^64, b∈ℝb the shared classifier parameters. Using layer normalization LNLN and the logistic sigmoid σ(z)=(1+e−z)−1σ(z)=(1+e^-z)^-1, the output is: iμ h_i^μ =LN(ϕu([ϕx(i)∥iμ])+ϕx(i)), =LN\! ( _u([ _x( x_i)\| M_i^μ])+ _x( x_i) ), (11) y^iμ y_i^μ =σ(⊤iμ+b). =σ( w h_i^μ+b). Let θ collect the trainable parameters. Then GATθ(Xt,ℰtμ)=(y^iμ)i∈tGAT_θ(X_t,E_t^μ)=( y_i^μ)_i _t (12) maps one graph view to one risk score per active agent; obstacle nodes are never scored. The self, message, and update maps are two-layer ReLU MLPs, and each attention map has one ReLU hidden layer and a single scalar output. In both instances, hidden layers have width 6464, each head outputs 1616 message features, and dropout is 0.100.10. The continuous model has 29,38129,381 trainable parameters; the grid model has 27,20527,205. Definition 7 (Collision warnings): Let c^i∈0,1 c_i∈\0,1\ be the checker decision for agent i: c^i=1 c_i=1 warns against its proposal, while c^i=0 c_i=0 allows it to pass. The deployed decisions are: c^icont c_i cont =y^iagt≥τagt∨y^iobs≥τobs, =1\ y_i agt≥ _ agt y_i obs≥ _ obs\, c^igrid c_i grid =y^iall≥τall∨y^iobs≥τobs. =1\ y_i all≥ _ all y_i obs≥ _ obs\. All views share one weight set, and their scores are tuned risks, not calibrated probabilities. The grid agent-only view is used only for diagnosis. iijjkkooi q_ij q_jk q_kstaticαji=.91 _ji=.91.06.06.03.03(a) black: proposed actions; blue: i’s attention,thickness == weight. j heads for i’sdestination and takes nearly all of iti M_i: weighted messages,four heads concatenatedi h_i: combined with ownstate ϕx(i) _x( x_i), residual ++ LNy^i=σ(⊤i+b) y_i=σ( w h_i+b)(b) summary ++ selfbecomes one risk score Fig. 3: CollisionGAT in one picture. (a) Black arrows are proposed actions; blue arcs are receiver i’s attention over its senders (Eq. 10), with thickness proportional to weight. Agent j proposes i’s destination and is emphasized by the illustrative attention weights. (b) The four-head summary is combined with the agent’s own state and mapped to one risk score (Eq. 11). IV-C Training and collision-detection results Discrete model. The grid model first learned elementary collision patterns from 40,000 primitive scenes and was then fine-tuned on 2,100 large-world joint-transition scenes. Its held-out test set contained 1,680 joint proposals and 8,364 agent decisions. It achieved precision, recall, and F1 of .9975, 1.0000, and .9987. Continuous model. The swept-motion model began with 2,800 primitive graphs. Task adaptation used 4,410 goal-directed graphs and 1,102 primitive replay graphs. A further 13,200 deployment-matched, difficult, and controller-generated graphs refined the frozen model. The frozen test contained 1,000 joint proposals from 100 new seed-disjoint world groups, producing 9,950 agent decisions. It achieved precision, recall, and F1 of .9127, .9996, and .9542. V Algorithm 1: Continuous Motion Algorithm 1 mounts the frozen CollisionGAT on a continuous waypoint controller. Before each episode, every agent independently runs deterministic A∗ on a four-connected lattice constructed from the known static-obstacle map [7]. The resulting path is greedily shortened into an ordered waypoint sequence i=(i,1,…,i,Ki=i)W_i=( w_i,1,…, w_i,K_i= g_i), with every retained straight segment checked for disc–obstacle clearance. During execution, the controller targets the current waypoint and advances to the next one when it is reached. At timestamp t, agent i samples ℓi,t∼U[.4,1.2] _i,t U[.4,1.2] and proposes a step toward its current waypoint i,t w_i,t: i=i+minℓi,t,‖i,t−i‖2i,t−i‖i,t−i‖2. q_i= p_i+ \! \ _i,t,\| w_i,t- p_i\|_2 \ w_i,t- p_i\| w_i,t- p_i\|_2. (13) The minimum shortens the final step so that the agent does not pass its waypoint. A moving proposal is unsafe if CollisionGAT warns it directly or if it can reach a warned waiting receiver in one step. Fair priority favors the longest-waiting agent. If their waiting times are equal, index priority rotates deterministically across timestamps. Algorithm 1 commits no GAT-positive moving proposal. We can showcase CollisionGAT because the obstacle-clear routes are planned independently, they do not coordinate interactions between agents. Algorithm 1 CollisionGAT-Guard (Continuous) 1: while t≠∅A_t≠ and the episode limit is unmet do 2: propose qiq_i using (13) for every active i; resolve shared targets by fair priority 3: assess q with the frozen GATθGAT_θ 4: while an agent is warned and fewer than five repairs were tried do 5: reduce one warned agent’s speed; reassess q 6: end while 7: hold remaining warned agents; restore one at a time only when reassessment gives no moving warning 8: if a moving warning remains then 9: ← q← p 10: end if 11: commit ← p← q synchronously; retire arrivals 12: end while Fig. 4: Full trajectories of two CollisionGAT mounts. Left: independently planned continuous waypoint routes create an agent–agent crossing conflict; CollisionGAT vetoes the proposal before controller repair. Right: obstacle-blind GATeD uses obstacle-attributed vetoes to discover and share wall cells before repairing D∗ Lite paths. Squares and stars denote starts and goals. VI Algorithm 2: GATeD The discrete Algorithm 2, GATeD, separates planning from checking. Agent i owns a complete four-connected grid graph Gi,tG_i,t with unit-cost available edges. The planner never sees the true obstacle set; CollisionGAT validates each joint proposal using only the local obstacle senders of Definition IV-A and returns typed warnings. Planning D∗(Gi,t,pi,gi)D^*(G_i,t,p_i,g_i) incrementally repairs agent i’s D∗ Lite state and returns its next minimum-cost cell, or pip_i at the goal or when no path exists. The shared discovered-wall set O starts empty. Private edge restrictions expire at the next timestamp, whereas shared wall restrictions persist. Vetoes The known-free set K initially contains every start and goal and gains every cell occupied after a committed move. An obstacle-attributed warning on qi∉q_i adds qiq_i to O and permanently removes that cell’s incident edges from every planner. Every other moving GAT warning or priority loss removes only (pi,qi)(p_i,q_i) from agent i’s graph until the next timestamp. If the same active agents return to the same positions without a new wall discovery, the controller cyclically selects one moving agent, temporarily blocks its next D∗ Lite edge, and validates the replanned proposal normally. Decisions Before learned warnings are acted upon, deterministic priority resolves duplicate destinations, swaps, and moves into cells held by waiting agents: a stationary occupant wins, otherwise the lowest-ID agent wins. Agents replan until a joint proposal is accepted or no repair remains; otherwise all agents wait for that timestamp. Exact grid predicates audit committed steps but never modify an action. Algorithm 2 GATeD (Discrete) 1: while t≠∅A_t≠ do 2: restore private edges; ←Falseaccepted 3: while not accepted and the repair budget remains do 4: qi←D∗(Gi,t,pi,gi)q_i ^*(G_i,t,p_i,g_i) for every active i; assess q with the frozen GATθGAT_θ 5: if duplicate destinations or swaps exist then 6: temporarily remove each losing agent’s edge (pi,qi)(p_i,q_i) from its own graph until the next timestamp 7: else if a moving proposal is warned then 8: permanently remove an obstacle-vetoed cell from every agent’s graph; otherwise remove only the warned agent’s edge (pi,qi)(p_i,q_i) until the next timestamp 9: else 10: commit ← p← q; update K; retire arrivals; ←Trueaccepted 11: end if 12: end while 13: if not accepted then 14: ← q← p 15: end if 16: end while The exact grid checks independently audit every committed step. Each planner keeps D∗ Lite’s incremental updates, while deterministic priority and temporary repairs coordinate the joint action. VII Evaluation Results We evaluate the continuous and discrete mounts on separate 100-world test suites. For the PNG maps, structural complexity is measured by the number of nodes in an occupancy quadtree, where fragmented free–obstacle boundaries require further subdivision [8]. The evaluated maps span three consecutive complexity bands: 15,00015,000–26,99926,999, 27,00027,000–38,99938,999, and 39,00039,000–50,99950,999 quadtree nodes. The continuous suite combines 50 generated 25×2525× 25 worlds containing square obstacles with 50 50×5050× 50 PNG maps from the lowest complexity band. Each episode contains 7–13 agents assigned to one to three goals. Of 994 agents, 846 arrived, and 66 of the 100 episodes completed. The remaining 34 episodes stalled with one or more agents short of their goals, showing that the waypoint controller does not guarantee completion. The left panel of Fig. 4 illustrates CollisionGAT vetoing an agent–agent crossing before controller repair. No GAT-positive moving proposal was committed, and the exact swept-motion audit recorded no executed collision. TABLE I: Controller outcomes on the two 100-world test suites. Setting # Worlds Agents arrived Collisions Discrete GATeD 100 995/995 0 Continuous CollisionGAT-Guard 100 846/994 0 The discrete suite contains 34 maps from the lowest complexity band, 33 from the middle band, and 33 from the highest band. It includes 49 50×5050× 50 and 51 100×100100× 100 maps, with 7–13 agents per episode and an even split between shared- and multiple-goal settings. Every episode completed, all 995 agents arrived, and no executed collision was recorded. Across 450,523 active-agent assessments, the obstacle view detected all 34,703 proposed wall entries with only one false positive. These vetoes accumulated 27,477 true obstacle-cell discoveries in the agents’ shared planning memory. VIII Discussion and Conclusion The main finding is that CollisionGAT can be mounted on substantially different controllers through the same propose–validate–update interface. The continuous waypoint controller and the obstacle-blind D∗ Lite planner use different state spaces, planning mechanisms, and responses to rejected actions, yet both provide CollisionGAT only with typed current and proposed states, and both recorded no audited executed collision across their respective 100-world evaluations. These results support portability at the interface level: a proposal-based controller that can repair, replan, or postpone a rejected action can use the checker without exposing its internal planning state. Because the two mounts use separately trained weights, this claim concerns the shared interface and architecture, not transfer of one checkpoint between motion domains. The arrival results distinguish collision screening from navigation completeness. The continuous controller delivered 846/994 agents, whereas GATeD delivered 995/995 (Table I); these are not accuracy comparisons because the controllers and test suites differ. The continuous mount may stall after conflicts or conservative warnings, while GATeD updates its planning graphs and retains discovered obstacle information. CollisionGAT screens proposed transitions; the controller remains responsible for progress, deadlock resolution, and arrival. Typed vetoes make the interface actionable: in GATeD, an obstacle-attributed veto adds the proposed cell to permanent shared wall memory, whereas an agent or priority conflict creates a private one-timestamp restriction, so CollisionGAT guides the controller’s response without receiving its planner state. Exact geometric predicates generate the training labels and independently audit every committed transition, so the results do not claim that CollisionGAT is more accurate than exact geometry or that it provides a formal safety certificate. Its value is its reusable relational interface: one frozen checker per motion domain converts a variable-size set of local agent and obstacle interactions into per-agent risk scores without receiving goals, routes, maps, or planner states. Geometry acts as teacher and evaluator, while CollisionGAT provides the learned validation layer mounted between proposal and execution. This collision-screening result is one instance of a broader idea: the same controller-agnostic validation interface can be mounted on other algorithms and extended to graph attention networks that veto actions violating two other important multi-agent requirements, connectivity and coverage. Acknowledgment OpenAI Codex and Anthropic Claude Code assisted with software refactoring, figure preparation, and language editing; the authors reviewed and verified all outputs. References [1] P. Veličković, G. Cucurull, A. Casanova, A. Romero, P. Liò, and Y. Bengio, “Graph attention networks,” in Proc. ICLR, 2018. [2] S. Brody, U. Alon, and E. Yahav, “How attentive are graph attention networks?” in Proc. ICLR, 2022. [3] N. Das and M. C. Yip, “Learning-based proxy collision detection for robot motion planning applications,” IEEE Trans. Robot., vol. 36, no. 4, p. 1096–1114, 2020. [4] C. Yu, H. Yu, and S. Gao, “Learning control admissibility models with graph neural networks for multi-agent navigation,” in Proc. 6th Conf. Robot Learning, PMLR, vol. 205, p. 934–945, 2023. [5] S. Zhang, O. So, K. Garg, and C. Fan, “GCBF+: A neural graph control barrier function framework for distributed safe multiagent control,” IEEE Trans. Robot., vol. 41, p. 1533–1552, 2025. [6] S. Koenig and M. Likhachev, “D* Lite,” in Proc. AAAI, 2002. [7] P. E. Hart, N. J. Nilsson, and B. Raphael, “A formal basis for the heuristic determination of minimum cost paths,” IEEE Trans. Syst. Sci. Cybern., vol. 4, no. 2, p. 100–107, 1968. [8] E. Meriaux, S. Wen, L.-R. Langevin, D. Precup, A. Loría, and G. Dudek, “On mobile ad hoc networks for coverage of partially observable worlds,” arXiv preprint arXiv:2512.09495, 2025.