1. Introduction
In modern emergency rescue operations, drone technology has become indispensable due to its rapid deployment, high mobility, and ability to operate in hazardous environments. However, conventional drone deployment strategies often treat communication, sensing, and computing as separate functions, leading to inefficient resource utilization and inadequate adaptation to sudden task arrivals. The heterogeneous nature of multi-source resources—such as computing power, bandwidth, and energy—further complicates the coordination problem. To address these challenges, we propose a novel method that jointly optimizes drone positions and multi-dimensional resources under an integrated sensing, communication, and computing (ISCC) architecture tailored for emergency scenarios.
Our work is motivated by the critical need to balance three core objectives: high communication data rate for real-time video backhaul, reliable coverage for first responders, and energy efficiency to extend operational endurance. By explicitly modeling uncertainties such as task burstiness, environmental dynamics, and link fluctuations, we construct a communication-sensing model that accurately captures the service capability. We then formulate a resource constraint mechanism that matches task requests with available heterogeneous resources. Finally, we define a comprehensive utility-maximization objective that weights the three conflicting goals. To solve the resulting complex optimization problem, we employ a multi-agent deep reinforcement learning framework that learns adaptive policies in a distributed manner. Experimental results demonstrate that our approach significantly outperforms static and single-objective baselines in terms of communication rate, coverage reliability, and energy consumption.
2. Related Work and Background
Drone technology has evolved rapidly in recent years. In disaster scenarios, drones are used as aerial base stations to restore connectivity when ground infrastructure is damaged. Early works focused on static placement or simple heuristic algorithms. For instance, Gupta et al. compared multi-objective metaheuristics for drone placement in post-disaster mesh networks. Other studies integrated mobile edge computing (MEC) with drone technology to reduce task latency. However, these approaches typically neglect the coupling between communication, sensing, and computing resources, leading to suboptimal performance in dynamic emergency environments.
The concept of ISCC—integrating sensing, communication, and computing on a unified platform—has emerged as a key enabler for next-generation networks. By sharing hardware and signals, ISCC improves spectrum and energy efficiency. Recent works in drone technology have explored ISCC for trajectory optimization and resource allocation, but most focus on generic scenarios without considering task urgency or real-time adaptation. Our method builds on these foundations by introducing a dynamic constraint-aware framework that explicitly accounts for task priorities and heterogeneous resource capabilities.
3. System Model and Problem Formulation
3.1 Uncertainty Factors
We consider a discrete-time system with time slots \(t = 1,2,\dots,T\). At each slot \(t\), a set of emergency tasks \(\mathcal{T}(t) = \{\tau\}\) arrives. Each task \(\tau\) is characterized by: estimated location \(\mathbf{p}_\tau\), a set of rescue personnel \(U_\tau = \{\mu_1,\mu_2,\dots\}\), real-time position of person \(u\) at time \(t\): \(\mathbf{q}_{\tau,u}(t)\), priority weight \(\omega_\tau \in (0,1]\), and maximum completion deadline \(D_\tau\).
There are \(K\) drones, indexed by \(k = 1,\dots,K\). Each drone \(k\) has maximum transmit power \(P_k^{\max}\), local computing capacity \(F_k^{\max}\), and a three-dimensional position \(\mathbf{p}_k(t)\). The initial positions are known but evolve over time.
The communication channel between drone \(k\) and the base station (backhaul) is:
$$
h_k(t) = \rho_k(t) \cdot \tilde{h}_k(t)
$$
where \(\rho_k(t)\) is the path loss dependent on distance, and \(\tilde{h}_k(t)\) represents small-scale fading. Similarly, the channel between drone \(k\) and rescue personnel \(u\) is:
$$
g_{k,u}(t) = \varsigma_{k,u}(t) \cdot \tilde{g}_{k,u}(t)
$$
with path loss factor \(\varsigma_{k,u}(t)\) and fading \(\tilde{g}_{k,u}(t)\).
3.2 Communication and Sensing Models
The backhaul data rate for drone \(k\) at slot \(t\) is given by:
$$
R_k^{\text{back}}(t) = B \log_2\left(1 + \frac{|\mathbf{w}_k^{\text{H}}(t) h_k(t)|^2}{I_k^{\text{back}}(t) + \sigma^2}\right)
$$
where \(B\) is the bandwidth, \(\mathbf{w}_k\) is the transmit precoder, \(I_k^{\text{back}}\) is interference, and \(\sigma^2\) is noise power.
The access link rate between drone \(k\) and user \(u\) is:
$$
R_{k,u}^{\text{acc}}(t) = B \log_2\left(1 + \frac{|\mathbf{v}_{k,u}^{\text{H}}(t) g_{k,u}(t)|^2}{I_{k,u}^{\text{acc}}(t) + \sigma^2}\right)
$$
with precoder \(\mathbf{v}_{k,u}\).
For sensing performance, we employ the Fisher information matrix. Define the steering vector \(\mathbf{a}_\tau(t)\) for task \(\tau\). The Fisher information \(J_\tau(t)\) is:
$$
J_\tau(t) = \frac{\dot{\mathbf{a}}_\tau^{\text{H}}(t) \mathbf{R}_x(t) \dot{\mathbf{a}}_\tau(t)}{\sigma_\tau^2}
$$
where \(\dot{\mathbf{a}}_\tau\) is the derivative, \(\mathbf{R}_x(t)\) is the total transmit covariance matrix, and \(\sigma_\tau^2\) is noise variance. The covariance matrix aggregates contributions from backhaul, access, and dedicated sensing signals:
$$
\mathbf{R}_x(t) = \sum_{k} \mathbb{E}[\mathbf{s}_k(t) \mathbf{s}_k^{\text{H}}(t)] + \sum_{k,u} \mathbb{E}[\mathbf{s}_{k,u}(t)\mathbf{s}_{k,u}^{\text{H}}(t)] + \sum_{\tau} \mathbb{E}[\mathbf{z}_\tau(t)\mathbf{z}_\tau^{\text{H}}(t)].
$$
A higher \(J_\tau(t)\) indicates better sensing accuracy for task \(\tau\).
3.3 Resource Constraints
We introduce the following constraints to ensure feasibility:
- Coverage constraint: Each task \(\tau\) must be covered by at least one drone:
$$
\sum_{k} a_{k,\tau}(t) \geq 1,\quad a_{k,\tau}(t) \in \{0,1\}.
$$ - Power constraint: Total transmit power of drone \(k\) at slot \(t\) cannot exceed its maximum:
$$
P_k^{\text{tx}}(t) \leq P_k^{\max}.
$$ - Access link quality: If drone \(k\) is assigned to task \(\tau\) and has arrived (indicated by \(a_{k,\tau}(t) \cdot \mathbb{1}_{\text{arrived}}\)), the signal-to-noise ratio (SNR) must be at least \(\gamma_{\text{rescue}}\):
$$
\frac{P_{k,u}^{\text{tx}}(t) |g_{k,u}(t)|^2}{\sigma^2} \geq \gamma_{\text{rescue}} \cdot a_{k,\tau}(t) \cdot \mathbb{1}_{\text{arrived}}.
$$ - Backhaul constraint: The backhaul SNR must exceed a threshold \(\gamma_{\min}\):
$$
\frac{P_k^{\text{back}}(t) |h_k(t)|^2}{I_k^{\text{back}}(t) + \sigma^2} \geq \gamma_{\min}.
$$ - Sensing accuracy constraint:
$$
J_\tau(t) \geq J_{\min} + \Delta_J \cdot \omega_\tau,
$$
where \(J_{\min}\) is a base threshold and \(\Delta_J\) scales with task priority. - Task processing delay constraint: The total delay for task \(\tau\) on drone \(k\) consists of local computation time, upload time, and edge computation time:
$$
T_{\tau}^{\text{proc}} = \frac{(1-\beta_{k,\tau}) D_\tau c_\tau}{F_k^{\text{loc}}} + \frac{\beta_{k,\tau} D_\tau}{R_k^{\text{back}}(t)} + \frac{\beta_{k,\tau} D_\tau c_\tau}{F^{\text{edge}}(t)} \leq D_\tau,
$$
where \(0 \leq \beta_{k,\tau} \leq 1\) is the offloading ratio, \(c_\tau\) is computation complexity, \(F_k^{\text{loc}}\) is local computing capacity, and \(F^{\text{edge}}(t)\) is edge server allocation. - Local computing capacity constraint:
$$
F_k^{\text{loc}}(t) \leq F_k^{\max}.
$$ - Total edge computing constraint:
$$
\sum_{k} F_k^{\text{edge}}(t) \leq F^{\text{total}}.
$$ - Trajectory dynamics constraint:
$$
\|\mathbf{p}_k(t+1) – \mathbf{p}_k(t)\| \leq v_{\max} \Delta t, \quad z_k(t) \geq z_{\min}.
$$
3.4 Objective Function
The goal is to maximize a weighted sum of communication rate, coverage reliability, and energy efficiency. Let \(R_{\text{total}}(t) = \sum_{\tau} \omega_\tau R_\tau(t)\) be the weighted sum rate, \(P_{\text{total}}(t) = \sum_k P_k(t)\) be total power consumption, and \(\text{Rel}(t)\) be the coverage reliability (e.g., fraction of tasks with satisfied links). The utility is:
$$
U(t) = \lambda_1 \frac{R_{\text{total}}(t)}{R_{\text{ref}}} + \lambda_2 \text{Rel}(t) + \lambda_3 \left(1 – \frac{P_{\text{total}}(t)}{E_{\text{total}}^{\max} / \Delta t}\right),
$$
where \(\lambda_1 + \lambda_2 + \lambda_3 = 1\) are dynamic weights adjusted based on mission phase. The overall optimization problem is to maximize \(\sum_{t=1}^T U(t)\) subject to all constraints.
4. Proposed Method: Multi-Agent Deep Reinforcement Learning
Given the high dimensionality and dynamic nature of the problem, we formulate it as a partially observable Markov game and solve it using a constraint-aware multi-agent deep reinforcement learning (MADRL) algorithm.
4.1 Problem Formulation as Dec-POMDP
We define the game \(\mathcal{G} = \langle \mathcal{N}, \mathcal{S}, \{\mathcal{O}_k\}, \{\mathcal{A}_k\}, \mathcal{R} \rangle\):
- \(\mathcal{N}\): set of agents (drones), \(|\mathcal{N}| = K\).
- \(\mathcal{S}\): global state space, containing positions, velocities, battery levels, task positions, priorities, channel gains, delays.
- \(\mathcal{O}_k\): local observation of agent \(k\), including own state, neighbor states within communication range, and current task list.
- \(\mathcal{A}_k\): action space of agent \(k\): \(\mathcal{A}_k = \{ \Delta \mathbf{p}_k, P_k^{\text{tx}}, \alpha_k, \mathbf{z}_k \}\) where \(\Delta \mathbf{p}_k\) is position adjustment, \(P_k^{\text{tx}}\) is transmit power, \(\alpha_k\) is compute offloading ratio, and \(\mathbf{z}_k \in \{0,1\}^M\) is task selection vector.
- \(\mathcal{R}\): global reward (sum of per-agent contributions) based on the utility function and constraint violation penalties.
4.2 Constraint Embedding
We embed hard constraints directly into the learning process:
- Power and dynamics constraints: After action selection, we perform feasibility checks and project actions onto the feasible set (e.g., clip power, scale displacement).
- Soft constraints (coverage, SNR, sensing, delay, local computing): Convert violation magnitude into negative reward signals. During training, the agents learn to avoid violations through trial and error. We use an adaptive penalty weight that increases when the violation rate is high.
- Global edge computing constraint: Incorporated into the critic network by adding a constraint satisfaction module that predicts the probability that the aggregate edge resource request does not exceed \(F^{\text{total}}\). This module adjusts the policy gradient update weight for states with high constraint satisfaction.
4.3 Reward Design
The reward at slot \(t\) is:
$$
r_t = r_t^{\text{util}} – \sum_{j} \eta_j \cdot \text{violation}_j(t),
$$
where \(r_t^{\text{util}}\) is the contribution to the objective function (scaled), and \(\text{violation}_j(t)\) are the magnitudes of constraint j violations, with penalty coefficients \(\eta_j\). The penalties are dynamic: they increase if the respective constraint is violated frequently.
4.4 Algorithm Implementation
We adopt a centralized training with decentralized execution (CTDE) paradigm. The critic network uses global state information to evaluate the joint policy, while each drone’s actor network outputs actions based on its local observation. We use Proximal Policy Optimization (PPO) with GAE for advantage estimation. The training procedure is summarized in Algorithm 1.
Algorithm 1: Multi-Agent PPO with Constraint Embedding
| Step | Operation |
|---|---|
| 1 | Initialize actor networks \(\{\pi_{\theta_k}\}_{k=1}^K\) and critic network \(V_\phi\) |
| 2 | Initialize replay buffer \(\mathcal{B}\) |
| 3 | for episode = 1 to MaxEpisodes do |
| 4 | Reset environment, obtain initial global state \(S_0\) and local observations \(o_{k,0}\) |
| 5 | for t = 0 to T-1 do |
| 6 | for each agent k do |
| 7 | Sample action \(a_{k,t} \sim \pi_{\theta_k}(\cdot | o_{k,t})\) |
| 8 | Apply feasibility projection (power clip, displacement clamp) |
| 9 | end for |
| 10 | Execute joint actions, environment updates to \(S_{t+1}\) |
| 11 | Compute reward \(r_t\) and obtain next observations \(o_{k,t+1}\) |
| 12 | Store transition \(( \{o_{k,t}\}, \{a_{k,t}\}, r_t, \{o_{k,t+1}\})\) in \(\mathcal{B}\) |
| 13 | end for |
| 14 | Sample mini-batch from \(\mathcal{B}\) |
| 15 | Compute advantages using GAE |
| 16 | Update critic by minimizing MSE between \(V_\phi(S)\) and returns |
| 17 | for each agent k do |
| 18 | Compute PPO surrogate objective with clipping |
| 19 | Update \(\theta_k\) via gradient ascent |
| 20 | end for |
| 21 | end for |
| 22 | Output trained policies \(\{\pi_{\theta_k}^*\}\) and critic \(V_\phi^*\) |
During execution, each drone only uses its local observation \(o_{k,t}\) and applies the deterministic policy (mean of Gaussian) to output actions. This enables real-time adaptation without centralized coordination.
5. Experimental Evaluation
5.1 Simulation Setup
We simulate an urban earthquake disaster scenario covering a 500 m × 500 m area with three blind zones (collapsed buildings). Ten rescue personnel are randomly distributed outside the blind zones. Three drones provide communication and sensing services. Drone parameters are listed in Table 1.
Table 1: Drone Configuration
| Parameter | Value |
|---|---|
| Battery capacity | 148 Wh (10,000 mAh / 14.8 V) |
| Max flight altitude | 200 m |
| Sub-6 GHz bandwidth | 100 MHz @ 3.5 GHz |
| mmWave bandwidth | 800 MHz @ 28 GHz |
| Communication module | 5G-A Cluster DRS |
| Sensing module | Thermal camera + 3D LiDAR |
| Edge computing resource | 4 cores, 16 GB RAM per drone |
| Onboard storage | 1 TB SSD |
We compare our proposed method against two baselines:
- Static resource allocation: Fixed transmit power and no offloading; drones hover at pre-planned positions.
- Function-split single-objective optimization: Only communication is optimized (max sum rate), with sensing and computing treated as separate subsystems.
All results are averaged over 10 independent runs with different random seeds. Training uses 2000 episodes; each episode lasts 300 time slots (1 slot = 1 second).
5.2 Communication Rate Comparison
Table 2 summarizes the key communication metrics.
Table 2: Communication Rate Performance
| Metric | Our Method | Static | Single-Objective |
|---|---|---|---|
| Average access rate (Mbps) | 150 | 65 | 86 |
| Peak access rate (Mbps) | 200 | 100 | 97 |
| Average backhaul rate (Mbps) | 1200 | 88 | 618 |
| End-to-end effective rate (Mbps) | 150 | 88 | 93 |
The results clearly show that our ISCC-based method achieves the highest rates in all categories, owing to dynamic resource allocation and integrated mmWave backhaul.
5.3 Delay and Reliability
End-to-end delay and packet delivery ratio are reported in Table 3.
Table 3: Delay and Reliability
| Metric | Our Method | Static | Single-Objective |
|---|---|---|---|
| Average end-to-end delay (ms) | 18.5 | 45.2 | 32.8 |
| High-priority task P95 delay (ms) | ≤25 | >60 | >60 |
| Packet delivery ratio (PDR) | ≥99.2% | 88.5% | – |
Our method keeps delays low even for high-priority tasks, thanks to adaptive offloading and trajectory optimization.
5.4 Coverage Reliability
Table 4 shows coverage-related metrics.
Table 4: Coverage Reliability Comparison
| Metric | Our Method | Static | Single-Objective |
|---|---|---|---|
| Link outage probability | 4.7% | 25.3% | 18.9% |
| Blind zone coverage rate | 95.2% | 60.3% | 71.5% |
| Average reliability probability | 95.3% | 74.7% | 78.6% |
Our method significantly reduces outages and covers almost all blind zones.
5.5 Energy Consumption
Table 5 breaks down energy usage.
Table 5: Energy Consumption (kJ)
| Component | Our Method | Static | Single-Objective |
|---|---|---|---|
| Total flight energy | 820 | 850 | 800 |
| Communication energy | 280 | 350 | 420 |
| Computation energy | 120 | 450 | 400 |
Our method reduces communication and computation energy by intelligent power control and edge offloading, while flight energy remains comparable.
5.6 Final Drone Positions
After the mission, the drones converged to the following optimal hovering points:
- UAV-1: (145 m, 160 m, 140 m) – covers Blind Zone 1 and assists right-side rescue team.
- UAV-2: (290 m, 260 m, 85 m) – low altitude over Blind Zone 2 for best sensing and access.
- UAV-3: (440 m, 340 m, 130 m) – covers Blind Zone 3 and left-side team.
This demonstrates the algorithm’s ability to achieve spatial self-organization based on task priorities and channel conditions.
6. Conclusion
In this work, we presented a novel drone deployment and resource co-optimization method under an integrated sensing, communication, and computing architecture for emergency scenarios. By modeling uncertainties, enforcing task-aware constraints, and maximizing a composite utility function, our method achieves superior performance in communication rate, coverage reliability, and energy efficiency compared to traditional static and single-objective approaches. The application of multi-agent deep reinforcement learning enables real-time adaptive decision-making in dynamic environments, highlighting the potential of drone technology in future intelligent emergency systems. Future work will focus on online learning with even tighter latency requirements and robust strategies for extreme weather conditions.