A drone that films terrain and radios it to a ground robot faces two problems that are usually solved separately. The first is what to send. Task-oriented "semantic" coding sends only the parts of an image that the robot's task actually needs, such as the narrow passages that decide a route. The second is where to fly. At very high radio frequencies, millimeter-wave and terahertz, a building between drone and robot does not merely weaken the signal. It cuts the link off almost completely. The drone's 3D path therefore decides when the link is up and when it is down. This proposal joins the two decisions. The drone should plan its flight so that the moments of guaranteed line of sight line up with the most important pieces of content, and it should not simply maximize the average number of bits delivered.
The reason this matters is that semantic payloads are lopsided. A few tokens carry most of the task value, and those same compact tokens are the most fragile when a link drops. A trajectory planned for average bit rate does not care whether its good link windows come early, late, briefly or reliably. A planner that knows which tokens matter does care. We expect this to change the flight path in exactly the regimes that matter in practice: dense buildings, tight battery budgets, and payloads larger than the link can reliably carry.
Core Claim
Under near-binary high-frequency blockage, the ordering of semantic importance creates a planning objective that is truly different from rate maximization. The best trajectory buys a high probability of line of sight during the time slots that carry high-importance content. It does not buy high average capacity. Because flight is slow and token scheduling is fast, the problem splits cleanly into alternating optimization: one fixes the other and solves, then they swap. One block is the 3D trajectory, solved with successive convex approximation over a smoothed blockage model. The other block is the assignment of tokens to time slots, which reduces to an exact linear program.
Let the image be split into token groups $k$ with task importance $u_k$ and size $s_k$. Let $\mathbf q$ be the discretized 3D trajectory with slots $n$, let $p_n(\mathbf q)$ be the line-of-sight probability in slot $n$, and let $R_n(\mathbf q)$ be the rate when the link is up. Let $x_{kn}\in[0,1]$ be the fraction of group $k$ sent in slot $n$; repeating a group across slots is allowed. We solve
$$ \max_{\mathbf q,\,x}\ \sum_k u_k\Big(1-\prod_n \big(1-p_n(\mathbf q)\big)^{x_{kn}}\Big)\quad \text{s.t.}\quad \sum_k s_k x_{kn}\le R_n(\mathbf q)\Delta,\ \ E_{\text{prop}}(\mathbf q)\le E_{\max}, $$
where $\Delta$ is the slot length and $E_{\text{prop}}$ is propulsion energy under an asymmetric climb, descent and turn model. This treats each slot's blockage as independent given the trajectory. The rate-maximizing alternative instead solves $\max_{\mathbf q}\sum_n p_n R_n$, and it is blind to which slots the critical tokens use. The two objectives coincide only when importance is uniform or when the reliable capacity exceeds the whole payload. That gives a sharp, testable prediction about when the joint design should help and when it should not.
Relation to Prior Work
The method comes from two blockage-aware lines of work. 3-D Positioning and Resource Allocation for Multi-UAV Base Stations Under Blockage-Aware Channel Model builds building polyhedrons from OpenStreetMap and smooths the sharp boundary between line of sight and blockage with a sigmoid, so gradient methods can optimize 3D positions inside a penalty-based double loop. Terahertz Communication Multi-UAV-Assisted Mobile Edge Computing System adds molecular absorption from HITRAN at 0.34 THz and reports penalty dual decomposition converging in 30–50 iterations. Both optimize static hovering positions for bit rate or delay. We reuse the smoothed blockage model and the absorption model, but we optimize a moving trajectory against a task-weighted delivery objective.
The semantic side motivates the work. Adaptive Semantic Communication for UAV/UGV Cooperative Path Planning (Path-SC) ranks image tokens by local betweenness centrality and drops the ones that do not matter for navigation. It sends about 54% less data than JPEG while keeping path-planning accuracy comparable, but the drone does not move and the channel is generic. We take its importance scores as our $u_k$.
Deep Joint Source-Channel Coding for Semantic Communications shows that learned joint coding degrades gracefully under additive noise. LaMI-GO and Vision Language Models for Massive MIMO Semantic Communication compress content into very sparse discrete anchors. None of these is tested against erasure. Graceful degradation under noise does not carry over to a link that is simply off, and this constrains our design: we must protect delivery itself, not rely on soft decoding.
Two papers do co-optimize semantics and flight. UAV-assisted Semantic Communication with Hybrid Action Reinforcement Learning uses 2D fixed-altitude flight and model-free PPO, and reports that a Triple PPO baseline timed out at eight users. Efficient Onboard Vision-Language Inference in UAV-Enabled Low-Altitude Economy Networks via LLM-Enhanced Optimization alternates image resolution and transmit power but plans the trajectory with reinforcement learning, and it models no blockage. Neither connects the timing of content importance to the timing of line-of-sight windows.
For energy, Multi-UAV path planning considering multiple energy consumptions via an improved bee foraging learning particle swarm optimization algorithm supplies the five-state propulsion model that we use as $E_{\text{prop}}$. That paper has no channel model at all.
Research Program
Three pieces will be built.
The first is a simulator. It places LandCover.ai orthophoto tiles as the sensed scene and draws urban building geometry from OpenStreetMap footprints and heights. It computes line of sight by ray–polyhedron intersection, using the sigmoid-smoothed version for optimization and the exact version for evaluation. It models links at 140 GHz and at 0.34 THz, the latter with HITRAN absorption. The sigmoid surrogate will be checked against ray-traced line of sight from the DeepMIMO O1 scenario at 140 GHz before it is trusted.
The second is the alternating solver.
With the trajectory fixed, the objective becomes concave in $x$ after a change to log-miss variables. The token-to-slot assignment is then an exact convex program. Its structure is simple: high-importance groups fill the high-probability slots first, then repeat into the next-best slots. With $x$ fixed, the trajectory step uses successive convex approximation on the sigmoid line-of-sight terms and on the propulsion energy, with speed and climb-rate limits.
The two blocks run on their natural time scales. The trajectory is re-solved every few seconds in receding-horizon fashion. Token assignment is re-solved every slot as importance scores and channel estimates arrive. This addresses the time-scale mismatch that flat alternating optimization ignores.
The solver will be started from both the rate-maximizing trajectory and a straight-line trajectory, so we can measure how sensitive the result is to initialization.
The third piece is the end task. The ground robot plans a route on the map it reconstructs from whatever tokens arrived. Success is measured on that route, not on image quality.
Baselines
The naive reference flies a straight path at fixed altitude and sends a JPEG image. Any proposal must beat it by a wide margin, or the problem is not real.
The strongest competitor is a separated design. Its trajectory maximizes expected rate, $\sum_n p_n R_n$, using the same smoothed blockage model and the same propulsion budget, extended from the static method of 3-D Positioning and Resource Allocation for Multi-UAV Base Stations Under Blockage-Aware Channel Model to a moving trajectory. On top of that fixed trajectory it runs Path-SC token selection with the same importance-greedy slot assignment and the same repetition we use. This baseline already has every ingredient except one: the trajectory step never sees importance. It isolates the paper's central claim and therefore doubles as the key ablation.
No numerical rate is reported on Lacuna for the static blockage-aware method. The paper states only a "significantly higher minimum achievable rate" than its baselines, so we will reimplement it. The semantic component's reported reference is Path-SC's roughly 54% data reduction over JPEG, and 43% over DeiT-SC, at comparable path-planning accuracy, on LandCover.ai. We will reproduce this on a clean channel before adding blockage.
A fixed-altitude 2D version of our own solver isolates what 3D flight adds. A hybrid-action PPO agent in the style of UAV-assisted Semantic Communication with Hybrid Action Reinforcement Learning gets the same simulator and the same wall-clock compute as our solver.
All methods share the energy budget, scenes, building maps, frequency settings and random seeds. Beating the separated baseline means a statistically significant gain in route validity at equal energy, or equal route validity at lower energy, across at least 50 OpenStreetMap city patches.
Why not just…?
Why not just maximize rate along the trajectory and let semantic coding handle the rest? This is the real threat. Our own objective predicts that the joint design adds nothing when the reliable capacity covers the whole payload or when importance is flat. The experiments therefore sweep the ratio of payload to reliable capacity and the skew of the importance scores, and report the regime boundary rather than one favorable point. If the separated baseline matches us across realistic densities, the honest conclusion is that slot scheduling is sufficient and joint trajectory design is unnecessary. That would still be a useful negative result.
Why not just fly high enough that buildings never block? Altitude does buy line of sight, but it also adds spreading loss, terahertz absorption that grows with distance as modeled in Terahertz Communication Multi-UAV-Assisted Mobile Edge Computing System, and the climb energy captured by the five-state model of Multi-UAV path planning considering multiple energy consumptions via an improved bee foraging learning particle swarm optimization algorithm. We include a tuned "climb to clearance altitude" baseline. We expect it to be competitive at 140 GHz and to lose at 0.34 THz, and we will report whichever outcome we find.
Why not just add retransmissions or forward error correction for the critical tokens? Blockage behind a building lasts seconds, not microseconds, so repeating a token inside a blocked stretch fails together with the original. Our repetition variable $x_{kn}$ already generalizes retransmission across slots. We will also run the separated baseline with an automatic-repeat protocol, which resends lost packets after a timeout. If it closes the gap, the benefit came from protection, not from flight planning.
Why not just train reinforcement learning end to end? The evidence in UAV-assisted Semantic Communication with Hybrid Action Reinforcement Learning is that hybrid-action PPO baselines become unstable as the problem grows; there, Triple PPO timed out at eight users. Decomposition methods, by contrast, converge in tens of iterations. Our PPO baseline at matched compute tests this directly in our setting.
Proposed Experiments
The primary metric is route validity: the fraction of ground-robot routes, planned on the received map, that avoid cells which are non-traversable in the ground-truth LandCover.ai map. Secondary metrics are path-length excess over the route planned on the true map, propulsion energy, delivered task utility, and solver time.
As a sanity anchor for map quality, the reported LandCover.ai segmentation reference is DeepLabv3+ at 85.56% mean intersection over union, from LandCover.ai. Roads score lowest at 68.74%, and roads are the class that matters for routing.
The decisive check is a two-dimensional sweep over building density and payload relative to reliable capacity, at both carrier frequencies. The prediction is that our gain over the separated baseline is near zero where payload is small or buildings sparse. It should grow where payload exceeds reliable capacity and buildings are dense. Its sign should be consistent with the change in where the line-of-sight probability is concentrated, which we can measure directly by comparing the line-of-sight profiles of the two trajectories.
A second check replaces Path-SC's importance scores with random scores at the same payload size. If our gain survives, the benefit comes from something other than semantics and the central claim fails.
Third, we compare the sigmoid surrogate with ray tracing on DeepMIMO O1, and we inject attitude jitter into beam pointing. This tests whether the planned line-of-sight windows survive realistic beam misalignment.
References
- 3-D Positioning and Resource Allocation for Multi-UAV Base Stations Under Blockage-Aware Channel Model
- Terahertz Communication Multi-UAV-Assisted Mobile Edge Computing System
- Adaptive Semantic Communication for UAV/UGV Cooperative Path Planning
- Deep Joint Source-Channel Coding for Semantic Communications
- LaMI-GO: Latent Mixture Integration for Goal-Oriented Communications Achieving High Spectrum Efficiency
- Vision Language Models for Massive MIMO Semantic Communication
- UAV-Assisted Semantic Communication with Hybrid Action Reinforcement Learning
- Efficient Onboard Vision-Language Inference in UAV-Enabled Low-Altitude Economy Networks via LLM-Enhanced Optimization
- Multi-UAV path planning considering multiple energy consumptions via an improved bee foraging learning particle swarm optimization algorithm
- LandCover.ai: Dataset for Automatic Mapping of Buildings, Woodlands, Water and Roads From Aerial Imagery