Communication-Aware Robot Execution for Cloud Inference under Spatially Heterogeneous Connectivity

arXiv:2606.31497 · cs.RO · Submitted 2026-06-30 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Today's paper: "Communication-Aware Robot Execution for Cloud Inference under Spatially Heterogeneous Connectivity".

Dev: Cloud-hosted foundation models enable robots to use semantic reasoning beyond onboard computational limits,

Rosa: First, who's behind it and why it matters.

Title and authors: Rosa: So let's start by looking at the title and who wrote this paper; it's "Communication-Aware Robot Execution for Cloud Inference under Spatially Heterogeneous Connectivity," authored by Fengkai Liu, Yuichi Ohsita, Masayuki Murata, and Hideyuki Shimonishi.

Dev: Those are some heavy hitters in the field; I wonder what their background brings to this specific problem involving spatial heterogeneity.

Taro: I've seen some work on decentralized consensus and communication efficiency papers like "NN-ETM: Enabling safe neural network-based event-triggering mechanisms for consensus problems," and I'm wondering if this paper draws from those ideas in terms of managing sparse communication.

Rosa: That’s a good connection, Taro; the idea of managing what data needs to be sent when, based on local conditions, seems central to their approach here.

Dev: I think the core implication is that they are moving away from just trying to find *any* place with a signal and instead focusing on finding a specific location where you can actually complete the entire cloud interaction cycle reliably.

Taro: If they successfully couple motion planning directly with communication constraints, it suggests we might see more robots capable of executing high-level semantic reasoning tasks in real-world, messy settings rather than just structured testbeds.

Rosa: I'm excited about that potential for deployment outside the lab; it feels like a step toward true autonomy in complex environments.

Dev: We’ll have to see if their estimated request–response window is accurate enough to keep the loop rate stable under these fluctuating conditions, though that seems like a major hurdle.

The paper's summary: Rosa: Now let's look at what the paper actually summarizes; it basically says that cloud-hosted foundation models let robots reason beyond their own hardware limits, but this execution gets shaky when connectivity isn't consistent spatially because the robot dictates *when* it needs a result, while the wireless environment dictates *where* it can send or receive data.

Dev: So they propose a framework that treats the next request point as a motion decision during ongoing execution, choosing that point to ensure enough communication quality for submission while still keeping track of where we are going for the response later.

Taro: It sounds like they are essentially solving a coordination problem between the robot's physical movement and the unpredictable wireless landscape simultaneously.

Rosa: Precisely; they introduce three main components: first, estimating that time window for a cloud cycle, second, optimizing exactly where to send the request point from, and third, planning the local path while keeping communication safety in mind after submission.

Dev: The way they define that request–response window as spanning from submission to retrieval—including transmission time and cloud inference—that seems like a very thorough way to account for all the delays involved.

Taro: I think the idea of defining a "robust request region" based on communication thresholds and a downstream retrieval margin, t ret(p), is smart because it doesn't just pick the closest signal; it picks one that guarantees success later.

Rosa: That focus on preserving progress within the finite support of the current primitive is key; they are making sure that even if we have to wait for a result, we haven't lost all our ground.

Dev: From an engineering standpoint, that suggests a more proactive approach than just reacting when the link fails; it’s about anticipating where you need to be before you send the next command.

The paper's improvements: Rosa: They suggest several specific improvements to make this framework more robust, like implementing a modular, hierarchical control architecture where the cloud handles high-level semantics and the robot manages the real-time geometric safety.

Dev: That architectural split sounds promising for managing complexity; it lets the cloud handle the heavy reasoning while we focus on keeping things stable locally.

Taro: I'm particularly interested in how they refine that request point optimization, using candidate filters to define a robust region R req t where points must satisfy communication thresholds and safety margins.

Rosa: That sequence of filters—forward region for progress preservation, downstream margin to ensure retrieval feasibility—that’s how they manage the spatial constraints effectively.

Dev: And I see them modifying the local path planning using Model Predictive Path Integral control, adding terms that incorporate dynamic communication costs into the objective function.

Taro: If they can successfully integrate that communication cost term, J comm(xi) =

zero S ret - M(r): S squared, it means the robot literally plans its path around signal degradation while still trying to get the task done.

Rosa: That level of integration between motion and communication is what really makes this approach different from just using a fixed connection map as a simple constraint.

Dev: However, I do want to mention one thing they flag as a limitation: the paper doesn't explicitly state how well this performs when the underlying cloud inference latency is highly variable or when the spatial connectivity map M(r) itself is very sparse and noisy.

Conclusion: Rosa: So, to wrap up, this paper introduces "Communication-Aware Robot Execution for Cloud Inference under Spatially Heterogeneous Connectivity," which tackles the fragility of using cloud models in unstable wireless environments by coupling motion planning with communication constraints.

Dev: They’ve shown that by estimating the request–response window and optimizing a robust request region R req t, robots can proactively move toward communication-favorable locations to submit requests while ensuring a feasible path exists for retrieving the result before the current task primitive expires.

Taro: The main implication is that this suggests we could deploy robots in truly dynamic, spatially varying settings where connectivity is unreliable, allowing them to perform complex reasoning tasks that were previously out of reach due to network instability.

Rosa: I think that's the big picture; it shifts the focus from just having a link to having a reliable *interaction cycle* under those conditions.

Dev: From an engineering view, the success hinges on their ability to accurately model that window estimation and keep their control loop rate stable despite these proactive movement decisions.

Taro: I think for future work, they should focus on validating this framework across a much wider range of connectivity scenarios than what's shown in their current experiments.

Rosa: That sounds like a solid direction; the next steps will be testing how resilient this co-design is when things get even messier.

Graduate School of Information Science and Technology, The University of Osaka

cs.RO

Submitted: 2026-06-30

Updated: 2026-10-02

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 76/100

The gist: Cloud-hosted foundation models enable robots to use semantic reasoning beyond onboard computational limits, but this execution becomes fragile under spatially heterogeneous connectivity because the

Key concepts

Request–Response Window Estimation
This estimates the total time needed for one full cloud interaction cycle, including sending data, waiting for the cloud to process it, receiving the answer, and accounting for uncertainty. It uses historical data (EWMA) and network models to predict this duration accurately.
Request Point Optimization
This selects a specific location where the robot should send its next request. The goal is to choose a point that offers good communication quality for submission but still allows the robot to reach a downstream retrieval spot later, preserving task progress.
Communication-Aware Local Path Planning
This component guides the robot's movement using Model Predictive Path Integral (MPPI) control. It plans paths toward either the optimized request point or the current target, incorporating constraints that prioritize communication quality during and after data submission.

Terminology

Summary

Cloud-hosted foundation models enable robots to use semantic reasoning beyond onboard computational limits, but this execution becomes fragile under spatially heterogeneous connectivity because the current primitive determines when the next result is needed, while wireless environment dictates where requests can be submitted and responses retrieved.

The gist: The proposed framework treats the next request point as a motion decision during ongoing primitive execution, selecting it to provide sufficient communication quality for cloud request submission while preserving progress within the finite support of the current primitive.

Framework Overview

The paper introduces a communication and motion co-design framework that couples high-level task scheduling with low-level robot navigation to coordinate local execution with cloud reasoning. This framework is built upon three main components:

  1. Request–Response Window Estimation, which estimates the stable connectivity duration required for one cloud interaction cycle by statistically predicting cloud inference latency from past observations together with network transmission time.

  2. Request Point Optimization, which computes a suitable request point such that the next cloud request can be submitted from a communication-sufficient region while still preserving a downstream opportunity to retrieve the returned result after passing that point.

  3. Communication-Aware Local Path Planning, which guides the robot toward the current active target, either the optimized request point or the task primitive target provided by cloud, while incorporating communication-aware soft constraints after submission and until the next cloud result is buffered locally.

Request–Response Window Estimation

This component estimates the time required for a next cloud cycle, denoted as the request–response window. This window spans from request submission to response retrieval and includes uplink transmission, cloud inference, downlink retrieval, and inference uncertainty. The predicted inference time is estimated using an Exponential Weighted Moving Average (EWMA) to smooth fluctuations:

Tbinfer(k + 1) = α Tobs infer(k) + (1 − α) Tbinfer(k)

The uplink transmission latency, Tup(r), is modeled by approximating the effective uplink SNR from the environmental communication map M(r). The expected goodput Rup(r) accounts for rate adaptation and retransmission overhead:

Rup(r) = Rlink(r) · (1 − ρ(r))γ

The final request–response window is estimated as:

Tbwindow(k + 1, r) = Tup(r) + Tdown(r) + Tbinfer(k + 1) + β σinfer(k + 1)

Request Point Optimization

This stage uses the estimated request–response window and the communication map to select where the next request should be submitted. The objective is to submit the request from a communication-favorable region while ensuring that the returned primitive can still be received as the robot continues toward the current waypoint. This is achieved through a sequence of candidate filters:

  1. Constructing a forward region (Rfwd t) that ensures Progress preserving and Bounded detour, requiring candidate points p to satisfy conditions like τmove(p, w) < τmove(r(t), w).

  2. Defining the downstream retrieval margin (∆tret(p)) as the largest slack between the time required to move from p to a downstream communication-feasible point q and the estimated request–response window for obtaining p(k + 1).

  3. Defining the robust request region (Rreq t) where points must satisfy communication thresholds and safety margins:

Rreq t = n p ∈ Rfwd t, M(p) ≥ Ssub, ∆ret t(p) ≥ εsafeo

The final request point p∗ is selected by minimizing the motion cost of reaching the current waypoint through the candidate point:

p∗ = argmin p∈Rreq t τmove(r(t), p) + τmove(p, w)

Communication-Aware Local Path Planning

The local planner generates motion while respecting robot dynamics and collision avoidance. It uses Model Predictive Path Integral (MPPI) control to solve a constrained optimization problem over a horizon H:

min u0:H−1 λnJnav (ξ; g (t)) + λuJctrl(ξ) + αt · Jca(ξ)

The active goal g(t) is set to the optimized request point p∗ before submission and returns to the current waypoint w after submission.

Improvements for AI systems

Here are specific improvements to AI systems based on the proposed framework, along with what those improved systems can achieve:

) Improved System Capabilities:

  1. Improved execution of robot-based Foundation Models in spatially heterogeneous and dynamic wireless environments by integrating motion planning directly with communication constraints.

  2. Enhanced robustness against network instability (fading, intermittent connectivity) during long-horizon cloud reasoning tasks by proactively selecting optimal request submission locations based on predicted communication windows and current primitive support.

  3. Reduced request failure rates (due to submission in poor connectivity regions) and minimized redundant/overlapped requests by using a communication-aware local planner that optimizes the robot's trajectory toward a request point rather than relying on fixed timing or naive polling.

) Specific Improvements:

  1. Implementation of the proposed framework into a modular, hierarchical control architecture (Cloud-Local Architecture) where the cloud provides high-level semantic guidance and the local controller handles real-time geometric grounding and safety assurance, as detailed in Section III-A.

  2. Integration of the Request–Response Window Estimation (Sec. III-C) using Exponential Weighted Moving Average (EWMA) to dynamically predict inference latency alongside uplink/downlink transmission times, allowing for an accurate calculation of the required communication budget for the next cycle:

Formula:

Tbwindow(k + 1, r) = Tup(r) + Tdown(r) + Tbinfer(k + 1) + β σinfer(k + 1).

  1. Development of the Request Point Optimization (Sec. III-D) module that utilizes a sequence of candidate filters (Forward Region, Downstream Retrieval Margin check, and Communication Feasibility check) to define a robust request region:

Condition for Robust Request Region:

Rreq t = n p ∈ Rfwd t where M(p) ≥ Ssub and ∆ret t(p) ≥ εsafe.

  1. Integration of the Communication-Aware Local Path Planning (Sec. III-E) using Model Predictive Path Integral (MPPI) control, modifying the local objective function to include dynamic communication costs:

Objective Function: min u0:H−1 λnJnav(ξ; g(t)) + λuJctrl(ξ) + αt · Jca(ξ).

Communication Cost Term: Jcomm(ξ) = max[0, Sret − M(rl)]∆S squared.

  1. Incorporation of the Arrival Time Penalty (Sec. III-E), which penalizes rollouts reaching the current waypoint before the response is buffered, ensuring that task progress is not prematurely halted:

Arrival Time Cost Term: Jarr(ξ) = max[0, Tbwindow(k + 1, p∗) − t − treq− (H∆t + τmove(rH, w)) 2].

  1. Application of the final selection criterion (Sec. III-D):

Request Point Selection: p∗ = argmin [τmove(r(t), p) + τmove(p, w)].

) Improved AI System Capabilities (What the improved system can do):

The resulting AI robot system can perform complex, long-horizon tasks in real-world environments with unstable connectivity by:

  1. Scanning its immediate environment (via the communication map) not just for obstacles, but for optimal communication staging areas that balance current task progress against future communication needs.

  2. Actively planning and executing a pre-emptive request maneuver—moving intentionally toward a specific, communication-favorable location to submit the next cloud query while ensuring it can still return to its primary task goal before the current instruction expires.

  3. Achieving significantly higher task success rates in challenging indoor settings (like those with corridors or signal drop-offs) compared to traditional methods, even when relying on a finite support of a cloud-provided semantic primitive.

  4. Operating reliably under conditions where network quality varies spatially, effectively treating the communication map as an integral part of the motion planning cost function rather than just a post-hoc constraint.

Sources

Related papers