Risk-Aware Joint Inference and Resource Control for Physical AI in 6G Networks

Amin Farajzadeh, Melike Erol-KantarciarXiv 2026 · arXiv

TL;DR

The paper introduces RACE (Risk-Aware Control for Edge inference), a two-timescale framework that optimizes 6G wireless resources for "Physical AI" (like robots) by accounting for the risk of incorrect AI decisions. By combining a slow-timescale Large Language Model (LLM) for task context with a fast-timescale reinforcement learning controller, RACE achieves an 11% improvement in efficiency while maintaining near-zero communication outages for critical tasks.

What Problem Does This Paper Address?

In the era of 6G, many robots and autonomous devices (Physical AI) will offload their complex AI processing to the "edge" of the network. This creates a tight coupling between wireless performance and physical safety.

Current wireless resource allocation typically uses fixed reliability targets, such as a constant Signal-to-Interference-plus-Noise Ratio (SINR). The authors argue this is fundamentally flawed for AI tasks:

  1. Overprovisioning: A robot navigating an empty hallway doesn't need ultra-reliable, high-energy communication for every frame.
  2. Underprovisioning: A robot approaching a person requires maximum reliability because the consequences of a misinterpretation are severe.

The gap lies in the lack of risk-awareness—the ability to dynamically adjust wireless protection based on both the likelihood of an AI error and the severity of that error's impact on the physical world.

Methods

The authors propose a two-timescale architecture called RACE to bridge high-level task intent with low-level radio control:

1. Slow-Timescale: The LLM "Contract"

Every few seconds (the task interval), an LLM analyzes a description of the robot's current task and environment. It proposes a "service contract" consisting of:

  • Physical Criticality: How much does an error matter right now?
  • Action Deadline: How fast must the inference result return?

Crucially, the LLM is not in the real-time control loop. Its proposals are checked by a deterministic validator to ensure they are feasible and safe before being passed to the fast-timescale controller.

2. Fast-Timescale: Risk-Aware Control

At the millisecond level (the wireless slot), a Hybrid Actor-Critic model (trained via PPO) makes decisions:

  • Uncertainty Estimation: A local predictor estimates the probability that the current sensory input will lead to an incorrect action.
  • Adaptive SINR Targets: RACE combines the LLM-derived criticality with the real-time uncertainty to set a custom SINR target. High risk = high SINR requirement.
  • Resource Control: The "Actor" selects the best AI model configuration (e.g., compression level) and transmit power.

3. The Restoration Layer

To handle the complex interference between multiple devices, the authors include a model-based power restoration layer. This layer analytically adjusts the power levels suggested by the AI to ensure they actually meet the required SINR targets, significantly improving the reliability of the system.

Datasets

The study evaluates RACE using a simulated interference-coupled uplink 6G network.

  • Simulation Environment: 6 devices offloading to a 12-antenna base station.
  • Profiles: The system uses pre-profiled data for different AI inference configurations, including encoded feature lengths, execution times, and expected action losses.
  • Training: The uncertainty estimation heads and RL controllers are trained on labeled offline data to map sensory inputs to failure probabilities.

Key Insights and Ideas

  • Risk = Probability × Consequence: This is the core "aha" moment. Most "task-oriented" communication only looks at the probability of error. By bringing in "consequence" (via the LLM), the network can prioritize energy for truly dangerous situations.
  • LLMs as Intent Interpreters: Instead of using LLMs to process raw data (which is slow and expensive), this paper uses them to interpret intent and set constraints. This allows the system to benefit from the LLM's reasoning without the latency penalty.
  • Separation of Concerns: The two-timescale approach ensures that the "messy" reasoning of an LLM is filtered through a deterministic validator before affecting the "hard" constraints of a wireless network.

Key Results

The authors compared RACE against several baselines, including "Fixed-high" (always high reliability) and "Uncertainty-only" (ignoring task criticality).

  • Efficiency Gain: RACE achieved an 11% improvement in a joint objective of accuracy, latency, and energy compared to the reliable "Fixed-high" baseline.
  • Reliability: As shown in Figure 2, RACE maintains near-zero constraint outages, whereas "Fixed-low" or "No-contract" methods suffer from massive failures.

![Fig.

  1. Average accuracy–latency–energy cost and constraint-outage rate.](/figures/art_408b659034294a32a067265603bff52f)
  2. The Value of the Contract: Figure 4 demonstrates that without the LLM-validated "contract" (the "No-contract" bar), the system's outage rate skyrockets to nearly 100%, proving that high-level task context is essential for edge AI.

![Fig.

  1. Slow-timescale contract ablation.](/figures/art_6b76d56996564e3b99e89c3b56cb69d0)
  2. Convergence: The training process for the hybrid controller is stable and converges quickly, as seen in the validation reward plots in Figure 1.

![Fig.

  1. Training convergence behavior.](/figures/art_ab95904544a048349c8118c45c53e1c6)

Limitations and Future Work

  • Out-of-Domain Tasks: The current model assumes the uncertainty head and performance profiles are valid for the current domain. If a robot enters a completely new environment, the system reverts to a "safe" fallback contract, which may be inefficient.
  • CSI Assumptions: The paper assumes perfect Channel State Information (CSI) at the base station. In highly dynamic 6G environments, CSI signaling overhead and errors could impact the restoration layer's effectiveness.
  • Future Directions: The authors suggest extending this to multi-cell scenarios and exploring how the LLM can more deeply integrate feedback from the physical environment to refine its criticality assessments.

Summary updated

Abstract

Wireless edge inference enables Physical-AI devices to execute edge-inferred actions using transmitted sensory inputs. Since the risk of incorrect actions vary across observations and tasks, the associated uplink transmission of sensory inputs requires different levels of wireless reliability. For instance, applying a fixed signal-to-interference-plus-noise ratio (SINR) requirement can overprovision low-risk tasks while underprovisioning high-risk ones. Therefore, we propose risk-aware control for edge inference (RACE), which maps predicted action-failure probability and task criticality to adaptive SINR targets and combines them with deadline-derived SINR requirements. A slow-timescale large language model (LLM) proposes task requirements for deterministic validation, while a fast hybrid actor selects inference configurations and transmit power levels, complemented by analytical receive beamforming and finite power restoration. Simulations show $11\%$ improvement in the average accuracy-latency-energy objective over fixed-high-SINR control, with near-zero constraint outage.
BibTeX
@misc{farajzadeh2026risk,
  author = {Amin Farajzadeh and Melike Erol-Kantarci},
  title = {{Risk-Aware Joint Inference and Resource Control for Physical AI in 6G Networks}},
  year = {2026},
  eprint = {2611.14696},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.14696},
}
Figure