PaperlayerView exact revision on arXiv

Source-linked AI summary

Safe Multi-Robot Coordination via VLM-LLM Reasoning and Reachability Analysis

Mohamed Dwedar, Ahmad Hafez, Alexander Jesser, Amr Alanwar

arXiv:2609.27816v1cs.ROcs.AIeess.SY

TL;DR

A camera-less robot must navigate using shared perception while ensuring the information is fresh, geometrically consistent, and safe for its own motion. The paper combines shared scene interpretation with robot-specific zonotope reachability checks that gate commands. In controlled indoor tests, it approved clear-path motion, redirected around dynamic human obstacles when a feasible corridor remained, and withheld commands when safety or freshness checks failed.

  • Problem

    A camera-less robot cannot observe its environment directly, so safe use of a partner’s perception requires validation against the receiving robot’s own state and motion constraints.

    Source: §I · p. 2

  • Method

    The Go2 supplies shared scene awareness while each robot’s own state parameterizes an independently propagated reachable tube used to verify its commands.

    Source: G. ZONOTOPE REPRESENTATION AND… · p. 9, §IX · p. 21

  • Results

    In controlled indoor tests, the pipeline approved clear-path motion, redirected around dynamic human obstacles when a feasible corridor remained, and withheld commands when reachability or freshness checks failed.

    Source: §V · p. 15, §IX · p. 21

  • Takeaways & Limitations

    Under the tested conditions, separating semantic reasoning from command authority and using robot-specific reachability gating supported safer heterogeneous-robot coordination.

    Source: §IX · p. 21

    The pipeline is not a formally safety-certified controller, and broader safety claims require diverse-scenario validation, worst-case analysis, hardware-in-the-loop testing, and long-duration trials.

    Source: §VIII · p. 20

Abstract

from arXiv · show

Safe coordination in heterogeneous machine-to-machine (M2M) robotic systems is challenging when robots differ in sensing capabilities, environmental awareness, and motion execution roles. This paper presents a centralized safety-aware M2M framework for cooperative goal-directed navigation in a heterogeneous mobile robot team comprising a vision-capable quadruped and a camera-less robotic vehicle. The objective is to guide both platforms toward a goal region while avoiding static and dynamic obstacles and preventing unsafe inter-robot interactions. Under the principle of shared perception, the vision-capable robot provides semantic environmental awareness through a centralized server over an MQTT broker, enabling the camera-less platform to navigate using this shared scene representation alongside its own odometry, IMU, and state feedback. A vision-language model (VLM) interprets the visual stream, and the extracted semantic data is mapped into conservative metric geometric constraints, including inflated obstacle sets, safe corridors, and goal regions. A large language model (LLM) proposes high-level task allocations, while physical command authority is restricted to a robot-specific zonotope reachability gate. This verification engine propagates independent reachable tubes to evaluate obstacle avoidance, safe-corridor containment, and inter-robot separation predicates before approving commands. Online experiments across clear-path and dynamic-obstacle scenarios show that the pipeline reliably approves safe motion, triggers conservative replanning or holding maneuvers upon constraint violation, and enforces a strict architectural separation between advisory semantic reasoning and formally verified motor execution.

I. INTRODUCTION

Safe coordination is difficult when a camera-less robot must act on shared perception that is fresh, geometrically consistent, and verified against its own state. The paper combines shared semantic-to-geometric perception with robot-specific reachability gates, keeping language-model proposals separate from command authority.

I. INTRODUCTION

  • Shared perception must be synchronized and checked against each robot’s own state before it can support safe motion.The camera-less platform otherwise lacks direct observation of obstacles, free space, and goals.Source: §I · p. 2
  • VLM and LLM outputs can be dynamically infeasible or miss collision geometry, so they are advisory proposals rather than substitutes for formal safety verification.Their latency and stochasticity also make them unsuitable for high-frequency low-level control.Source: §I · p. 2

A. METHOD AT A GLANCE

  • The pipeline combines MQTT communication, Go2-derived shared perception, conservative geometric constraints, symbolic task proposals, and reachability checks for a camera-less SVEA.SVEA also contributes synchronized proprioceptive state, and its commands are independently verified.Source: §I · p. 2
  • The five-gate cascade checks timestamp consistency, VLM schema, geometric consistency, robot-specific reachability, and command freshness before publication.A failed gate yields a conservative HOLD or STOP; only the reachability engine can approve a command.Source: A. METHOD AT A GLANCE · pp. 2–3, Figure 1 · p. 3
  • The semantic-to-geometric module turns structured scene descriptions into conservative obstacle sets, unsafe-region inflation, safe corridors, and goal regions.Its safety margins account for localization, perception, communication, and conservative uncertainty components.Source: A. METHOD AT A GLANCE · p. 3

B. WHAT IS ADOPTED, WHAT IS ADAPTED, AND WHAT IS NEW

  • The paper adapts reachability supervision to robot-indexed verification and decomposes obstacle inflation using perception confidence and reprojection error.Its new couplings transfer one robot’s visual constraints to another, test inter-robot separation on paired tubes, and make freshness depend on end-to-end latency.Source: B. WHAT IS ADOPTED, WHAT IS ADAPTED,… · pp. 3–4

II. RELATED WORK

Prior work treats communication, shared perception, foundation-model reasoning, and reachability safety largely in separate contexts. This paper positions its contribution as their integration for asymmetric sensing, with shared constraints evaluated against robot-specific reachable sets.

A. M2M COMMUNICATION AND MULTI-ROBOT COORDINATION

  • M2M communication work primarily studies transport and reliability, while this pipeline confines MQTT to carrying telemetry, camera streams, and approved commands.The broker performs no planning, reasoning, or verification, and communication failures yield conservative command rejection.Source: A. M2M COMMUNICATION AND MULTI-ROBOT… · p. 4

B. HETEROGENEOUS ROBOTS AND SHARED PERCEPTION

  • Existing shared-perception systems often omit synchronization, coordinate transformation, or formal verification for the receiving robot.Here, SVEA contributes state feedback and its commands are verified using its own kinematic state and reachable set.Source: B. HETEROGENEOUS ROBOTS AND SHARED… · p. 4

C. VLM AND LLM REASONING FOR ROBOTICS

  • Foundation models support embodied reasoning, but cannot reliably enforce kinodynamic constraints or guarantee collision-free motion across unfamiliar scenes.The pipeline extends a single-robot reachability filter to a heterogeneous two-robot team with asymmetric sensing.Source: C. VLM AND LLM REASONING FOR ROBOTICS · p. 4

D. REACHABILITY ANALYSIS AND ZONOTOPE-BASED SAFETY VERIFICATION

  • Zonotopes enable closed-form propagation under affine maps and Minkowski sums, supporting computationally tractable online safety-predicate evaluation.Prior reachability work does not address shared-perception asymmetry or language-model-generated proposals in heterogeneous multi-robot coordination.Source: D. REACHABILITY ANALYSIS AND… · pp. 4–5

E. RESEARCH GAP AND POSITIONING

  • The paper integrates shared VLM scene interpretation, LLM-supported symbolic coordination, MQTT, and zonotope verification into one online safety-filtered pipeline.Against twelve prior systems, it reports 0/12 for both constraint transfer to a platform without exteroceptive sensing and paired-tube inter-robot separation; the proposed pipeline satisfies 6/6 capability axes.Source: E. RESEARCH GAP AND POSITIONING · p. 5

III. METHODOLOGY

  • Both robots send state data through MQTT, while the Go2 also streams camera frames for semantic interpretation and geometric constraint construction.The broker is asynchronous transport only; it does not plan, reason, or verify commands.Source: §III · p. 5

A. NOVELTY DELINEATION WITHIN THE METHODOLOGY

  • The methodological novelty lies in coupling shared constraints to distinct robot tubes, paired-tube separation, and a freshness condition tied to measured pipeline latency.The source of constraints can differ from the robot whose own state and input bounds parameterize verification.Source: A. NOVELTY DELINEATION WITHIN THE… · p. 5

B. ROBOT STATE AND STREAM SYNCHRONIZATION

The server synchronizes robot streams and reduces each robot’s state to the planar configuration needed for safety checks. It converts validated visual semantics into conservative metric constraints, with calibration accuracy limiting the reliability of projected obstacle boundaries.

B. ROBOT STATE AND STREAM SYNCHRONIZATION

  • The reduced planar state retains position and heading for occupancy checks, while velocity is represented through input and disturbance bounds rather than as an explicit state.This is an intentional short-horizon safety-envelope model, not a full mechanical reproduction of either platform.Source: B. ROBOT STATE AND STREAM… · pp. 5–6
  • Packets with timestamps outside the synchronization tolerance are discarded, and the server issues HOLD or STOP rather than reasoning over mismatched streams.This prevents a stale camera frame from being paired with a current robot state for safety verification.Source: B. ROBOT STATE AND STREAM… · p. 6

C. SEMANTIC SCENE INTERPRETATION

  • Each validated VLM description represents an entity by its semantic class, image-space polygon, confidence score, and timestamp; malformed or inconsistent JSON is discarded.Confidence is retained as a field because it cannot alone determine the spatial safety buffer; reprojection quality also matters.Source: C. SEMANTIC SCENE INTERPRETATION · p. 6

D. SEMANTIC-TO-GEOMETRIC CONVERSION

  • Camera calibration and floor-plane projection map image regions into metric obstacle footprints, safe corridors, and goal regions; systematic calibration error can bias obstacle boundaries beyond the inflation model.The corridor is reduced by the robot footprint to represent center-point reachable space, and goal regions come from semantic cues such as open-door directions or waypoint markers.Source: D. SEMANTIC-TO-GEOMETRIC CONVERSION · p. 7
  • Obstacle inflation combines robot footprint, localization, perception, delay, and margin buffers, with perception inflation depending on both VLM confidence and reprojection error.A scalar buffer is a deliberate simplification that trades potential corridor width for robustness and computational tractability.Source: D. SEMANTIC-TO-GEOMETRIC CONVERSION · p. 7
  • A path must remain inside the safe corridor and outside inflated obstacles, but satisfying these geometric conditions alone does not approve a command.The full reachability gate must still be passed before publication.Source: D. SEMANTIC-TO-GEOMETRIC CONVERSION · p. 8

E. LLM SYMBOLIC TASK REASONING

The LLM reasons over validated scene information and both robots’ current states to propose a symbolic task-level action. That proposal guides candidate-control selection but has no authority to issue a motor command; the reachability gate can reject it.

E. LLM SYMBOLIC TASK REASONING

  • Qwen3 uses the validated scene representation and both robots’ current states to propose a symbolic task-level action.Its output is a candidate input to the reachability gate, not a continuous motor command.Source: E. LLM SYMBOLIC TASK REASONING · p. 8
  • The symbolic modes are PROCEED, AVOID_AND_PROCEED, REPLAN, HOLD, and STOP, covering supervised low-speed goal-directed navigation behaviors.They distinguish forward motion, obstacle avoidance, path replacement, pausing, and terminal unsafe or emergency directives.Source: E. LLM SYMBOLIC TASK REASONING · p. 8
  • The gate rejects any LLM-derived candidate control that violates a geometric or reachability constraint, so the LLM exercises no command authority.The LLM provides high-level structural guidance that reduces the search space for candidate controls.Source: E. LLM SYMBOLIC TASK REASONING · p. 8

F. ROBOT DYNAMICS AND LINEARIZATION

The framework uses a reduced planar unicycle model and local affine linearization to bound short-horizon robot occupancy. It propagates independent zonotope reachable tubes for the two platforms, representing bounded inputs and disturbances while keeping computation tractable.

F. ROBOT DYNAMICS AND LINEARIZATION

  • Both platforms use a reduced planar unicycle approximation for short-horizon occupancy bounds, with approximation error treated as a bounded disturbance.The model is an intentional safety-envelope reduction, not a claim that the robots share identical dynamics.Source: F. ROBOT DYNAMICS AND LINEARIZATION · p. 8
  • Local linearization yields an affine model whose residual is included in the disturbance set, so propagated tubes account for approximation error.Affine maps of zonotopes remain zonotopes, enabling closed-form propagation.Source: F. ROBOT DYNAMICS AND LINEARIZATION · p. 8

G. ZONOTOPE REPRESENTATION AND REACHABLE TUBE PROPAGATION

  • Zonotope generators represent independent uncertainty directions, and affine maps and Minkowski sums preserve the representation for tractable set-membership checks.Order reduction bounds generator count as sums accumulate generators.Source: G. ZONOTOPE REPRESENTATION AND… · p. 8
  • The server propagates Go2 and SVEA reachable tubes independently because their states and kinematic envelopes differ, and issues no averaged or shared command.Each tube represents planar configurations possible over the horizon under admissible inputs and bounded disturbances.Source: G. ZONOTOPE REPRESENTATION AND… · pp. 8–9

H. SAFETY PREDICATES AND DECISION RULE

Commands are approved only when reachable tubes avoid inflated obstacles, remain within safe corridors, and stay separated from the other robot. Approved commands are robot-specific and freshness-checked before execution; online validation assesses the complete pipeline under controlled laboratory conditions.

H. SAFETY PREDICATES AND DECISION RULE

  • A command is approved only if each reachable tube avoids inflated obstacles, stays inside the safe corridor, and does not overlap the other robot’s tube.All three predicates must hold simultaneously.Source: H. SAFETY PREDICATES AND DECISION RULE · p. 9
  • A tube that exits the corridor without intersecting an obstacle triggers holding and replanning, while obstacle intersection or predicted inter-robot collision triggers STOP.PROCEED and AVOID+PROCEED are both SAFE outcomes; their distinction is encoded in the candidate control derived from the LLM proposal.Source: H. SAFETY PREDICATES AND DECISION RULE · p. 9

I. COMMAND PUBLICATION AND FRESHNESS

  • Each robot receives a separately computed command packet, and execution is permitted only while the packet remains within its validity horizon.Stale packets are discarded because the scene or robot state may have changed since the server’s snapshot.Source: §I · p. 9

IV. EXPERIMENTAL SETUP AND ONLINE VALIDATION

  • Online validation evaluates whether synchronized sensing, interpretation, geometric conversion, verification, and publication operate as a causally ordered loop under controlled laboratory conditions.The stated goal is complete online pipeline operation, not offline replay or isolated module testing.Source: §IV · p. 9

A. PLATFORM CONFIGURATION

The experiments pair a camera-equipped Go2 with a camera-less SVEA, using shared perception and independently verified robot-specific commands. Across the illustrated cases, reachability checks approve feasible motion and reject or redirect proposals that violate obstacle or corridor constraints.

A. PLATFORM CONFIGURATION

  • The Go2 supplies the live camera stream while camera-less SVEA contributes its own synchronized state and receives commands verified against its own pose and reachable set.The robots’ distinct trajectories and workspace positions motivate independent command checks rather than one shared command.Source: A. PLATFORM CONFIGURATION · p. 10, E. EXPERIMENTAL FIGURES · p. 14

B. IMPLEMENTATION STACK

  • The centralized stack uses VLM scene interpretation and LLM task proposals, but grants physical command authorization exclusively to the zonotope reachability gate.MQTT carries telemetry and approved commands; it performs no perception, planning, reasoning, verification, or control.Source: B. IMPLEMENTATION STACK · p. 10

C. ONLINE DECISION LOOP

  • Each online cycle synchronizes inputs, validates VLM output, converts scene descriptions into metric constraints, and evaluates symbolic proposals using independent robot-specific reachable sets.Failures at synchronization, schema validation, geometric conversion, reachability, or freshness checks lead to conservative rejection.Source: C. ONLINE DECISION LOOP · p. 10

D. EXPERIMENTAL SCENARIOS AND SAFETY PRECAUTIONS

  • Trials covered static clear-path and moving-human scenarios under supervised laboratory conditions, with speed limits, workspace boundaries, and emergency-stop access.The authors state these precautions are not formal safety certification or a guarantee outside tested conditions.Source: D. EXPERIMENTAL SCENARIOS AND SAFETY… · p. 10

E. EXPERIMENTAL FIGURES

  • In the clear-path case, both robots’ reachable tubes satisfy obstacle avoidance, corridor containment, and inter-robot separation, so the gate publishes robot-specific PROCEED commands.The camera-derived scene is converted into geometric constraints before motion approval.Source: E. EXPERIMENTAL FIGURES · p. 12
  • When a candidate reachable tube extends beyond the verified traversable corridor, the gate rejects it and assigns REPLAN even if obstacle clearance holds.The command is withheld until a new candidate path remains inside the safe corridor and satisfies obstacle separation.Source: E. EXPERIMENTAL FIGURES · p. 12
  • During dynamic-obstacle interaction, the gate issues REPLAN or HOLD when a reachable tube intersects the inflated human-obstacle region or leaves the safe corridor.A path that clears the human footprint can still be rejected if it exits the verified corridor.Source: E. EXPERIMENTAL FIGURES · p. 14
  • Obstacle inflation accounts for semantic confidence and geometric projection uncertainty, so visually confident but geometrically uncertain detections remain conservatively bounded before reachability checks.Reprojection error, calibration uncertainty, partial visibility, and perspective can require a larger safety radius even at high confidence.Source: E. EXPERIMENTAL FIGURES · pp. 14–15

F. REPRODUCIBILITY

The paper records inputs, intermediate outputs, decisions, timing, and configuration so that online decision cycles can be reconstructed offline.

F. REPRODUCIBILITY

  • Each decision cycle logs sensor and communication timestamps, model outputs, geometric sets, reachable tubes, decisions, commands, latencies, and failure events.The records include frame identifiers and timestamps for camera, odometry, IMU, and both robots’ state data.Source: F. REPRODUCIBILITY · p. 15
  • Configuration files preserve robot initial states, goal definitions, obstacle settings, safety buffers, prediction horizon, communication settings, model identifiers, and stochastic seeds.The passage says the full configuration record supports reconstruction of each online cycle.Source: F. REPRODUCIBILITY · p. 15
  • The logging structure is intended to enable offline reconstruction of every online decision cycle for analysis and verification.Source: F. REPRODUCIBILITY · p. 15

V. RESULTS AND DISCUSSION

The evaluation examines gate behavior across laboratory scenarios, including nominal acceptance in clear-path navigation. In that case, the shared scene supports feasible motion and the reachability gate approves separate commands for both robots.

V. RESULTS AND DISCUSSION

  • The authors report outcomes consistent with expected pipeline behavior under tested laboratory conditions, including rejection events at pipeline gates.Source: §V · p. 15

A. TEST MATRIX

  • The evaluation recorded 1,953 decision cycles over 65.1 seconds at 30 Hz, verifying 17,577 candidate arcs across nine curvature candidates per cycle.Each arc was propagated for 15 zonotope steps at 0.1-second intervals, covering a 1.5-second horizon.Source: A. TEST MATRIX · p. 15

B. CLEAR-PATH SCENARIO

  • In the clear-path scenario, both robots’ reachable tubes satisfy obstacle avoidance, corridor containment, and inter-robot separation, and the gate publishes approved robot-specific PROCEED commands.The candidate path lies within the safe corridor and outside the inflated obstacle set.Source: B. CLEAR-PATH SCENARIO · p. 15

C. DYNAMIC-OBSTACLE SCENARIO

The dynamic-obstacle evaluation shows how the gate redirects motion around a detected human and approves commands only when reachability constraints hold. Across scenarios, approvals track geometry, while perception statistics support using obstacle geometry—not confidence alone—to characterize risk.

C. DYNAMIC-OBSTACLE SCENARIO

  • When a human entered, the initial path intersected the inflated unsafe region; AVOID+PROCEED redirected it, with approval only when all three predicates held and REPLAN or HOLD otherwise.The gate checked obstacle avoidance, corridor containment, and inter-robot separation before approval.Source: C. DYNAMIC-OBSTACLE SCENARIO · pp. 15–16

D. QUANTITATIVE GATE BEHAVIOR

  • 56.4% of cycles were approved in clear-path S1 versus 0.0% at stairwell S2; the corridor predicate rejected all 361 S2 cycles despite no obstacle intersection.Obstacle avoidance alone would have approved forward motion toward the descending staircase.Source: D. QUANTITATIVE GATE BEHAVIOR · p. 16, Table 4 · p. 17

E. DECISION-CONDITIONED STATISTICS

F. PERCEPTION STATISTICS AND VALIDITY OF THE UNCERTAINTY MODEL

  • Obstacle geometry explained 36.8% of cycle-risk variation versus 9.1% for confidence; 89 detections without resolvable confidence were retained as maximum-uncertainty cases.This supports decomposing the safety inflation rather than scaling a single term by confidence alone.Source: F. PERCEPTION STATISTICS AND VALIDITY… · pp. 16–17
  • The system detected a mean of 0.50 objects per cycle, and empty-detection cycles were not treated as free space because the corridor predicate remained active.S2 rejected every candidate despite its empty obstacle set.Source: F. PERCEPTION STATISTICS AND VALIDITY… · p. 17

G. REJECTION CASES AND CONSERVATIVE BEHAVIOR

The safety gate overrides proposals when reachability, synchronization, or perception checks fail, while each robot receives an independent evaluation. The reported comparisons and deployment discussion bound the evidence to controlled laboratory scenarios and identify conditions needing further work.

G. REJECTION CASES AND CONSERVATIVE BEHAVIOR

  • Obstacle-tube intersection, corridor exit, inter-robot separation violations, synchronization failures, and schema-inconsistent VLM output triggered conservative rejection without sending commands to execution.The gate overrode the LLM proposal in every rejection case.Source: G. REJECTION CASES AND CONSERVATIVE… · p. 17

H. MULTI-ROBOT INDEPENDENCE AND SVEA ACTIVE

  • Go2 and SVEA received different commands at each timestep, with SVEA evaluated using its own synchronized state and reachable set rather than Go2’s state.Both robots actively contributed state data to the cooperative task.Source: H. MULTI-ROBOT INDEPENDENCE AND SVEA… · p. 17

I. COMPARISON WITH THE STATE OF THE ART

  • Unlike language-model planners without a runtime gate, this system can override a symbolic proposal; at stairwell S2, the corridor predicate prevented motion despite semantically traversable floor.The cited comparison notes that those planners do not expose a kinodynamic verification stage.Source: §I · p. 18
  • Reachability planners without semantic input require an externally supplied obstacle set, whereas semantic safe-corridor constraints let this gate reject the obstacle-free stairwell case.The cited comparison identifies the hazard as non-traversability rather than a geometric obstacle.Source: §I · p. 18
  • The comparison with single-robot filter [28] rests on shared constraint transfer to the camera-less robot and a predicate coupling two reachable tubes.The cited single-robot formulation has neither mechanism.Source: §I · p. 18
  • A direct numerical benchmark was not attempted because baselines differ in platforms, sensing, and success criteria, and none reports the matching camera-less approval measure.The authors consider a numerical comparison inadvisable under these conditions.Source: §I · p. 18

J. GENERALIZABILITY TO REAL-WORLD DEPLOYMENT

  • The authors expect the advisory-semantics/verified-authority separation and five-gate cascade to transfer across platforms with bounded inputs and disturbance envelopes.They also expect weak confidence–risk correspondence to recur with single-camera ground-plane obstacle geometry.Source: J. GENERALIZABILITY TO REAL-WORLD… · p. 18
  • Denser scenes may lower approval rates, while outdoor terrain, variable lighting, and centralized-server failure remain deployment boundaries not resolved by this validation.The 1.5 s horizon is described as adequate at 0.09 m s−1 mean speed but needing growth with velocity, trading tube size against conservatism.Source: J. GENERALIZABILITY TO REAL-WORLD… · p. 18

K. DECISION LOGIC SUMMARY

Approved commands require all reachability predicates and freshness checks to pass; latency determines how long commands remain valid. The evaluation describes a representative latency budget and baseline/ablation protocol, but not instrumented latency results or quantitative module-removal comparisons.

K. DECISION LOGIC SUMMARY

  • PROCEED and AVOID+PROCEED are issued only when all three reachability predicates and the command-freshness condition are satisfied.The decision table maps reachability conditions and active predicates to decisions and command status.Source: K. DECISION LOGIC SUMMARY · p. 18, Table 7 · p. 19

VI. LATENCY AND RUNTIME ANALYSIS

  • The latency budget sums MQTT, synchronization, VLM, geometry, LLM, reachability, and command stages; its values are representative design-time estimates, not measured results.Instrumented latency benchmarking over extended trials is identified as future work.Source: §VI · p. 18
  • VLM inference contributes 750 ms, LLM reasoning 300 ms, and reachability propagation 100 ms for two robots under the stated configuration.MQTT transport, synchronization, geometric conversion, and command publication collectively contribute approximately 50 ms under laboratory network conditions.Source: §VI · p. 18
  • With an adopted total budget of approximately 1200 ms, command validity must cover total latency plus margin; stale commands are rejected even if reachability predicates pass.The experiments set Tvalid to 1.5 s for clear-path and 2.0 s for dynamic-obstacle scenarios.Source: §VI · pp. 18–19
  • The implementation targets supervisory low-speed navigation and safety-filtered command updates, not high-frequency stabilization, which remains on robot controllers.On-robot controllers handle stabilization and emergency braking at their native frequencies.Source: §VI · p. 19

VII. BASELINE AND ABLATION EVALUATION

  • Baseline conditions remove modules such as geometric verification or shared perception to isolate their intended roles, but the protocol reports no quantitative module-removal differences.Systematic quantitative evaluation across these conditions is future work.Source: §VII · p. 19, Table 9 · p. 20
  • Ablations vary safety-buffer size, semantic uncertainty, communication delay, and prediction horizon to examine their expected effects on acceptance, freshness rejection, and corridor containment.These are evaluation conditions and intended contributions, not reported numerical outcomes.Source: §VII · pp. 19–20

VIII. LIMITATIONS

The validation is limited to controlled, low-speed indoor settings and depends on reliable perception and calibration. The conclusion reports reachability-gated coordination results but does not claim a formally safety-certified controller.

VIII. LIMITATIONS

  • Validation covered two robots in controlled indoor labs, with static boundaries, one moving human, and relatively uncluttered scenes.Outdoor settings, variable lighting, occluded views, multiple simultaneous dynamic obstacles, and teams larger than two were not evaluated.Source: §VIII · p. 20
  • VLM scene interpretation and camera calibration constrain the safety predicates, and calibration bias cannot be fully compensated by obstacle inflation.VLM failures, partial detections, or domain shift can produce incorrect obstacle sets or safe corridors.Source: §VIII · p. 20
  • The centralized server concentrates perception, reasoning, and verification, making it a potential bottleneck and single point of failure.A server failure would interrupt the command pipeline for all robots.Source: §VIII · p. 20
  • VLM and LLM inference limits updates to a supervisory timescale suited to low-speed local navigation, not high-speed or highly dynamic tasks.Faster inference, asynchronous pipelines, or smaller domain-specific models could reduce this limitation.Source: §VIII · p. 20
  • The reduced unicycle model estimates planar occupancy over a short horizon but omits full SVEA nonholonomic and Go2 legged dynamics.A more complete model could reduce conservatism and improve command acceptance in more demanding maneuvers.Source: §VIII · p. 20
  • The pipeline is not a formally safety-certified controller, and broader safety claims require systematic validation, worst-case analysis, hardware-in-the-loop testing, and long-duration trials.Extended measured-latency benchmarking is also required to report runtime performance rather than design-time estimates.Source: §VIII · pp. 20–21

IX. CONCLUSION

  • Across 1,953 decision cycles, the gate approved motion in 39.8% of cycles and withheld it in 60.2%.The validation covered clear-path and dynamic-human-obstacle scenarios.Source: §IX · p. 21
  • The centralized design shares Go2 perception with camera-less SVEA and reserves command authority for reachability checks; among 13 systems, it alone provides constraint transfer and paired-tube separation.The gate also withholds commands that fail a predicate or freshness condition.Source: §IX · p. 21
Loading 2609.27816v1…