LUCID: An Agentic AI Framework on Digital-Twin in the Loop for QoS-Guaranteeing Robotic Control

Hyeonsu Lyu; Minwoo Kim; Sehyun Ryu; Hyun Jong Yang

Source-linked AI summary

LUCID: An Agentic AI Framework on Digital-Twin in the Loop for QoS-Guaranteeing Robotic Control

Hyeonsu Lyu, Minwoo Kim, Sehyun Ryu, Hyun Jong Yang

arXiv:2608.28437v1Aug 28, 2026eess.SYcs.RO

TL;DR

Cloud robotics lacks a responsive way to adapt coupled trajectory and radio-resource decisions as wireless conditions and operator intents change. LUCID dynamically configures TP–RRM schemas in a DITL pipeline, and experiments show robust adaptation, including operation of 15 robots after reducing the required throughput from 15 Mbps to 8 Mbps.

Problem
Changing robot positions, transmissions, and scenes make trajectory-dependent wireless feasibility dynamic, while fixed TP–RRM formulations and costly scene evaluation limit responsive orchestration.
Method
LUCID uses an LLM to configure bounded TP–RRM schemas, SimBridge to create wireless-ready digital twins, and deterministic path-planning and spectral-radius RRM validation to search feasible states.
Results
LUCID adapts to changing workload requirements, allowing all 15 robots to operate safely after reducing the required throughput from 15 Mbps to 8 Mbps.
Takeaways & Limitations
Dynamic, intent-driven TP–RRM schema orchestration is feasible in a large-scale DITL testbed.

Abstract

from arXiv · show

Cloud robotics relies on the timely uplink of high-volume sensing streams, yet dynamic environments continually shift the feasible combinations of trajectories, active-robot count, and per-robot QoS. Because existing approaches formulate trajectory planning (TP) and radio resource management (RRM) as a single fixed optimization problem, they cannot reconfigure these coupled decisions as conditions evolve, resulting in transient QoS violations. However, evolving operator intents change which quantities-such as the active-robot count and per-robot QoS-are fixed, optimized, or relaxed. Furthermore, the computational cost of evaluating trajectory-dependent wireless conflicts has made it difficult to build large-scale Digital-Twin-in-the-Loop (DITL) testbeds responsive enough for such dynamic orchestration. We present LUCID, an LLM-agent--orchestrated, uplink-aware cloud-robotics pipeline that moves TP--RRM from solving a fixed formulation to dynamically orchestrating optimization problem schemas within a DITL environment. Driven by the operator's high-level intent, LUCID treats the TP--RRM formulation as a bounded template whose variables, objectives, and constraints are dynamically configured, while SimBridge enables repeated ray-tracing evaluation by converting large-scale robotics scenes into wireless-ready DTs. By integrating collision-free path planning with a spectral-radius RRM validator, LUCID identifies wireless bottlenecks and restructures the problem schema on the fly to efficiently find the verified feasible state. Experiments confirm that LUCID robustly adapts to changing intents, active-robot counts, and scenes, while a multimodal surrogate model, FastConfigNet, reduces planning latency.

I. INTRODUCTION

Cloud robotics must jointly manage robot trajectories and uplink resources because changing positions, interference, blockages, and operator priorities alter feasible operating points. LUCID addresses this by dynamically configuring TP–RRM schemas in a DITL pipeline and supporting repeated wireless evaluation.

Motivation: Cloud robotics must identify supportable combinations of active-robot count, per-robot uplink QoS, and trajectories before deployment.Shared, spatially varying wireless links remain vulnerable to transient blockages and multipath fading.
Coupled challenges: Independent trajectory planning can create physical conflicts or concentrate collision-free robots in radio-congested regions.TP and RRM therefore cannot be treated as independent decisions.
Coupled challenges: Moving robots and simultaneous transmissions continually change interference, blockage, and multipath conditions across space and time.Localized channel degradations can produce transient QoS violations that conventional statistical channel models may smooth out.
DITL challenge: Large robotics scenes lack wireless material descriptions and contain geometric detail that makes repeated ray-tracing evaluation computationally expensive.This overhead limits their use in responsive feasibility-search loops.
Intent conditioning: Operator priorities change which TP–RRM variables are fixed, optimized, or relaxed, so feasibility search must be intent-conditioned and structured.LUCID translates each search step into a verifiable TP–RRM problem.
Proposed framework: LUCID combines SimBridge, LLM-guided schema configuration, and FastConfigNet to search feasible operating points defined by active AMRs and per-AMR QoS.SimBridge creates wireless-ready digital twins, while FastConfigNet screens candidate configurations before costly evaluation.

II. INTENT-CONDITIONED TP–RRM SYSTEM MODEL

LUCID maps operator intent and a digital-twin environment to instances of a bounded TP–RRM template. The planner generates candidate trajectories, while deterministic RRM validation tests wireless feasibility and returns diagnostics for revising configurations or paths.

Workflow: The DITL workflow translates an operator request into a candidate operating-point configuration that instantiates the TP–RRM template.The agent revises configurations as planning and validation modules evaluate each instance.
System model: The environment combines a materialized, geometry-reduced SimBridge scene, base-station layout, and ray-tracing results for trajectory-dependent channel queries.The configuration contains deployment, motion, obstacle, timing, QoS, resource, and noise parameters.
System model: The centralized edge server jointly plans robot motion and uplink scheduling over discrete time slots subject to motion and resource constraints.Robot positions, destinations, speeds, obstacle positions, rate requirements, bandwidth, power, and noise are represented in the model.
Optimization template: The template minimizes a weighted sum of travel distance and uplink transmit power while enforcing per-robot rate requirements and resource limits.The weight λ controls the tradeoff between motion and transmission energy terms.
Optimization template: The solver generates candidate trajectories and performs slot-wise RRM feasibility evaluation over association, bandwidth, and power variables.The formulation includes binary trajectory and association decisions together with nonconvex rate constraints.
Intent conditioning: Intent determines which configuration fields remain fixed and which vary across operating points, including relaxation of active-robot count or rate requirements.For an infeasible requested count, the agent follows the specified relaxation priority and reevaluates a revised configuration.

III. WIRELESS-AWARE TP–RRM SOLVER

The solver couples collision-free trajectory planning with per-slot wireless feasibility validation, using conflict diagnostics to revise trajectories until all slots are feasible or the search fails.

RRM Feasibility Validation: The RRM validator tests candidate trajectories against DT-derived channel gains and returns feasible allocations or wireless conflict diagnostics to the trajectory planner.It evaluates each time slot after robot positions and predicted channel gains are fixed.
RRM Feasibility Validation: For fixed association and bandwidth allocation, spectral-radius conditions determine whether finite power control exists and whether the power cap is satisfied.When ρ(T(t)S(t)) < 1, the least feasible power vector is obtained from the nonnegative inverse; feasibility also requires the resulting power to respect ρmax.
Frequency Allocation and User Association: The validator refines user association through strongest-link seeding, random restarts, and one-user reassignment moves that preserve feasibility while reducing the evaluated power objective.Bandwidth allocations are recomputed at the two affected base stations after each reassignment.
Feasibility Diagnostics: It distinguishes interference bursts from power-cap violations and localizes each failure to a culprit robot using spectral-radius tests and diagnostic rules.The complete trajectory is solver-verified only when every time slot returns a feasible allocation.
Wireless-Conflict-Aware Multi-Robot Planning: The navigation graph is built by thresholding a DT depth map, sampling nodes with clearance, rejecting obstacle-crossing edges, and removing disconnected components.Standard CBS then generates joint trajectories on the resulting collision-free graph.
Wireless-Conflict-Aware Multi-Robot Planning: Wireless CBS augments geometric collision detection with RRM validation, converting failed slots and culprit robots into robot-specific space–time constraints.This represents physical and wireless conflicts through the same CBS constraint interface while retaining the core search logic.

IV. DATA LAKE AND FASTCONFIGNET

LUCID archives evaluated configurations and uses FastConfigNet to prioritize likely-feasible operating points before expensive TP–RRM solver validation.

Data Lake: LUCID records operator intents, instantiated configurations, DT environments, candidate trajectories, optimized RRM variables, and solver outcomes in a data lake.These records support learning-based preconditioning of repeated trajectory-dependent validation.
FastConfigNet: FastConfigNet combines global scene and system-configuration features in a prediction head that estimates candidate failure probability.Training uses class-weighted binary cross-entropy, stratified splits, and early stopping based on validation loss.
FastConfigNet: During intent-driven search, FastConfigNet forwards a high-QoS configuration whose predicted feasibility probability exceeds a threshold to the TP–RRM solver.For a fixed 10-AMR request, the search terminates after the validator confirms the selected configuration as feasible.

V. SIMBRIDGE: FROM ROBOTICS SCENES TO WIRELESS-READY DTS

SimBridge converts rendering-oriented robotics scenes into wireless-ready digital twins by adding electromagnetic material information and reducing geometry while preserving propagation-relevant structure. Its mesh simplification trades path-gain fidelity for lower scene size and evaluation time.

Scene preparation: SimBridge transforms robotics scenes with geometry, asset instances, and visual or physical metadata into material-aware, geometry-reduced scenes for wireless propagation evaluation.It addresses missing radio-material identifiers and the high cost of repeated propagation evaluation.
Scene preparation: A vision-language mapping assigns each asset an ITU electromagnetic material identifier from metadata and a thumbnail.The model output is canonicalized to a material identifier and propagated to associated scene instances.
Geometry simplification: Material-preserving simplification aggregates transformed geometry by material, triangulates meshes, clusters vertices by spatial voxels and normals, and recomputes area-weighted normals.This preserves coarse surface geometry and material assignments while supporting accurate ray incidence and reflection angles.
Geometry simplification: The full warehouse shrinks from 68,314,408 vertices, 89,481,477 faces, and 1.3 GB to 12,196,481 vertices, 12,066,933 faces, and 350.6 MB.The reduced representation substantially lowers scene storage and geometric complexity.
Fidelity–cost trade-off: At 10% mesh retention, path-gain RMSE is about 1.1 dB and runtime is 0.64 s, versus 1.07 s for the 100% reference.Lower retention reduces representation size and evaluation time while increasing error relative to the unsimplified mesh.

VI. NUMERICAL EXPERIMENTS

Experiments evaluate LUCID’s schema-selection reliability, adaptation across changing intents and environments, end-to-end performance, and computational efficiency. Results show that stronger LLM backends and wireless-aware orchestration support reliable reconfiguration, while FastConfigNet reduces validator overhead and preserves deterministic verification.

Intent-to-configuration reliability: Claude Sonnet 5 and Gemini 3.6 Flash achieve schema-selection accuracies of 94% and 91%, followed by GPT-4.1 mini at 83%.GPT-4.1 nano and Llama 3.1 8B reach 62% and 39%, respectively.
Intent-to-configuration reliability: Raw validator rejection rates are 0–1% for the three strongest backends but 60–72% for smaller backends, while semantic escape remains 6–16% across models.The stronger backends are more reliable under prompt injection and ambiguous or conflicting requests, though accuracy is not perfect.
Multi-phase adaptation: LUCID lowers the required throughput from 15 Mbps to a feasible 8 Mbps when adding five robots, allowing all 15 robots to operate safely.The agent prioritizes active-robot count over individual data rates and reconstructs the optimization schema accordingly.
Multi-phase adaptation: After environmental changes, LUCID continues serving 15 AMRs by further lowering the QoS threshold, whereas baselines without scene updates fail to satisfy QoS requirements.SimBridge converts the changed scene and the DT runtime recomputes ray-tracing channel maps before reevaluation.
End-to-end comparison: Under moderate load, LUCID’s QoS-weighted job throughput is approximately 1.9× higher than Separated TP/RRM, while its goal-reaching ratio remains 1.0 versus about half.Average travel distance and transmit power remain nearly identical between methods.
End-to-end comparison: Under heavy load, LUCID preserves a goal-reaching ratio of 1.0 while Separated TP/RRM approaches zero completion, producing more-than-one-order-of-magnitude throughput separation.The comparison also reports a similarly large separation in QoS-weighted service throughput.
Runtime and throughput: FastConfigNet screens candidate trajectories before deterministic RRM validation, reducing expensive validator calls as link-aware planning overhead grows with AMR count.Geometric planners remain fast because they handle only physical space-time conflicts, whereas wireless-aware planners also screen throughput constraints.
Runtime and throughput: As AMR count increases, wireless-aware methods achieve higher per-robot throughput than geometric planners by avoiding interference and power-cap violations.All algorithms reach nearly the maximum target rate with 4 users, while throughput decreases as congestion grows.

VII. CONCLUSION

LUCID reframes cloud-robotics TP–RRM as dynamic orchestration of predefined optimization schemas configured by operator intent and scene context. The results demonstrate that this intent-driven approach is feasible in a large-scale DITL testbed despite prior fixed-formulation and testbed limitations.

Conclusion: LUCID dynamically configures TP–RRM variables, objectives, and constraints within a predefined template as operator intent and scene context evolve.The framework combines LLM schema configuration with deterministic wireless-aware planning and RRM validation.
Conclusion: Dynamic, intent-driven TP–RRM schema orchestration is feasible in a large-scale DITL testbed, overcoming fixed-formulation and testbed limitations.