# Drone Path Planning using AI Topic radar · 3,370 documents · 2021-01-01–2026-10-27 · updated 2026-10-28 Canonical: https://slate.greyb.com/radar/drone-path-planning-using-ai/innovations ## In one minute ### Every route in this field comes out of the same three-part chain Almost nothing in this field plans a path in one step. A search layer fixes the order in which stops, structures or targets are visited. A trained policy flies the legs in between and handles what the search could not foresee. A check layer decides whether the result is allowed to run. The pattern holds across delivery, inspection, agriculture and defence work. Nanjing University of Aeronautics and Astronautics puts a genetic algorithm on top and reinforcement learning underneath [1]. A genetic algorithm is a search that breeds candidate routes and keeps the better ones. Reinforcement learning trains a policy by trial and error against a reward. Beihang's earlier template separates a logistics stage that fixes the delivery order from a flight stage that generates the trajectory [2]. The dependency runs both ways. A 2024 Hefei filing has reinforcement learning tune the genetic algorithm's own crossover and mutation rates [3]. 1. [CN121212410A](/?paper=CN121212410A&utm_source=radar&utm_medium=dashboard) 2. [CN113283827A](/?paper=CN113283827A&utm_source=radar&utm_medium=dashboard) 3. [CN118747568A](/?paper=CN118747568A&utm_source=radar&utm_medium=dashboard) ### The market record is programmes, service revenue and test infrastructure, and very little of it is a purchase of planning The demand that reaches a buyer is demand for flying. ZenaTech guided to C13.4m to C14.8m for the third quarter of 2026 on 27% organic growth in its drone-as-a-service operations, having reported C$7.73m for the first nine months of 2025 and a 1,225% year-over-year quarter. The rest of the field's public activity is grant-funded or demonstrative: a European project ran GPS-free, beyond-line-of-sight navigation for logistics drones in cities to the end of 2023 and closed without a product, and ACSL and Preferred Networks reached demonstration stage in September 2026 on generating routes from mission requirements under a Japanese national programme. Where a planner was released as an actual artefact, it was given away: MIT and Penn open-sourced a motion planner benchmarked at about 90% of the runtime of state-of-the-art methods and roughly 15% faster to destination, built explicitly to replace commercial solvers costing six figures. ### Ten months of 2026 already match the whole of 2025 Publication ran 121 documents in 2021, 185 in 2022, 286 in 2023, 519 in 2024, 1,143 in 2025 and 1,116 through 28 October 2026. The steep part of that curve was 2023 to 2025, when output roughly quadrupled in two years. The 2026 figure covers ten months, so on a monthly basis the field is publishing about a sixth faster than it did last year. The tripling has stopped and the volume has held. ### The ordering step became a sequence model Where the search used to be a genetic algorithm, the newer filings learn the order instead. Tsinghua encodes coordinates and demand as vectors and lets an attention mechanism weight each unvisited node, trained by policy gradient. Attention lets the model learn which candidate stop matters most at each step. Anhui University couples a task decoder and a path decoder in one Transformer, so sequence and trajectory are chosen together. Liaoning Technical University folds charging stops into the same graph, splitting nodes into delivery, pickup and charging. ### Where autonomy does sell as a distinct product, it sells as a layer over someone else's airframe Auterion's order to fit terminal guidance to 50,000 Ukrainian Shrike drones is worth roughly 103m and prices the added autonomy near 2,060 on an airframe that costs about $400. That ratio, roughly five to one in favour of the software, is the closest thing in this record to a price for navigation separated from the aircraft. ### No area is declining, and the two smallest grew fastest Comparing the ten months of 2026 with the twelve months of 2025, delivery sits at 352 against 350, energy at 198 against 201, infrastructure at 124 against 117 and agriculture at 67 against 66. Annualised, all four exceed last year. Defence at 230 against 251 and public safety at 134 against 148 are the two whose raw totals fall short, and they fall short only because the comparison year was complete. ### The map moved onto the aircraft, mid-flight The older pipeline surveyed the asset, built a model, then planned over the model. The newest filings plan from what the sensors see during the flight. A substation planner rebuilds its model while airborne and absorbs equipment updates without a full replan. A line planner infers components it could not observe, then generates passes to check the inferred areas. A fault-aware line planner recomputes safe boundaries from the condition it detects on the structure it is circling. A solar planner reroutes on live obstacle data. ### The gate that actually opened is regulatory, and it is contested Executive Order 14307, signed 6 June 2025, directed the FAA to fast-track routine beyond-visual-line-of-sight rulemaking and to use AI tools to speed waiver processing. A performance-based proposed rule followed in August 2025, covering low-altitude BVLOS operations and the third-party traffic management services that support them. In July 2026 the FAA signed the environmental assessment and finding of no significant impact for commercial package delivery under Part 135, and states are now challenging that national review, which leaves the clearance in litigation rather than settled practice. ## Overview Method claims are dense and cheap to file. Demonstrations, orders and flight hours are scarce. Among inventors the split is 1,885 academic documents against 1,438 corporate ones, and the corporate block is not drone makers. It is State Grid and its provincial units, Guangdong Power Grid, Huaneng, China Mobile, China Electronics Technology Group. The organisations that already own wires, towers and spectrum are writing their own planning code instead of buying it. ## Research ### The objective moved from distance to energy Distance minimisation has given way to sortie economics. Charging stops became nodes in the graph. One filing jointly optimises flight paths and charging stops across a fleet, and another routes drones between charging stations instead of returning home. Nest placement is optimised in its own right, with drones assigned to zones sized by endurance. ### Constraints became regions, so a check layer enforces them Wind started as a cost penalty on a route and ended as a zone the aircraft may not enter. Positioning error became an obstacle modelled as a probability cloud. A digital-twin airspace controller now scores real-time risk and triggers automatic warnings or flight bans, and an identity system builds a temporary virtual fence around each verified drone before granting it a path. ### Sensing is fused before any route is drawn, so the plan is a forecast The most complete public-safety version aligns lidar, thermal and gas readings into one picture, has a language model write a plain situation report, and plans against combined energy, time and safety costs. A tunnel swarm holds position on ultra-wideband radio where satellite positioning fails, and arbitrates between disagreeing sensors with fuzzy logic. A drone and a robot dog share one synchronised map. ### A traffic authority now sits above the aircraft The newest layer lives off the aircraft. An identity system predicts each drone's behaviour, checks the claim against a blockchain record, then builds a fence and an air corridor around the verified drone, adjusting corridors to avoid conflicts. A companion filing predicts flight intentions with recurrent networks and long short-term memory, flags anomalous behaviour, and sends a geofenced path back over a quantum-encrypted link. ### The evidence of any of it working comes from simulators the applicants built The delivery work computes every performance figure on the applicant's own map. One university claims a 65% cost cut against planning the parts separately, another a 42% cut in travel time. No field trial or independent benchmark accompanies either. ## Market ### Where the hard half is being paid for, it is the public purse Because operators do not publish their navigation stacks, the record of who funds the difficult work is the grant list. A UK research programme, ATRACT, is funded to build a drone system that helps frontline medics prioritise battlefield casualties, and a UK innovation-funded project completed field testing of secure, repeatable BVLOS operations in safety-critical airspace. Peer-reviewed work models multi-vehicle drone delivery where the 2024 Guangdong floods cut the roads, and schedules water trucks alongside firefighting drones to compensate for payload limits. ### Every sizing figure is a forecast, and none of them isolates the planner A syndicated report puts AI path planning at 2.40bn in 2025 rising to 19.91bn by 2034 at 26.5% a year. Its own segmentation runs across automotive and mobility, so the drone share cannot be separated out, and the source is vendor research. A second forecast puts the commercial drone market above $90bn by 2036, with services the largest segment and hardware the fastest growing. India's industry body measures hardware at 54% of industry value in a $1.31bn market for 2025-26, with software and drone-as-a-service the fastest-growing segment. ### What would make this a market rather than a literature No tender in this record has a buyer specifying behaviour. The procurement evidence names platforms, payloads and docks; the standards work in the field is going the other way, mapping machine output onto risk thresholds and ratings that already exist rather than certifying the search itself. ## Strategy ### Companies now write nearly half of it, and they are utilities and telecoms Corporate documents went from 38 in 2021 to 517 in 2026, against 83 and 596 for academic ones. The corporate share of the yearly total rose from about a third to about 46 percent. That is the clearest structural change in the period. ### The biggest cluster in the field is radio work filed by telecoms universities The largest single cluster holds 370 documents and is about radio, not parcels. It is trajectory and task offloading for drones used as edge-computing nodes, where the path decides how well the aircraft serves devices on the ground. Chongqing University of Posts and Telecommunications ranks joint fifth in the whole field with 43 documents, and the University of Electronic Science and Technology of China holds 33, Xidian 30 and Beijing University of Posts and Telecommunications 20. These are telecommunications departments, and they have treated the flying path as a variable in a network optimisation problem. ### China is the filing office for nearly all of it, and India is the only other sustained contributor Chinese offices account for 1,045 of 1,135 delivery documents, 638 of 714 in defence, 575 of 592 in energy, 354 of 399 in public safety, 299 of 319 in infrastructure and 159 of 179 in agriculture. India is second in every area where anyone else appears, at 30 defence documents, 22 public safety, 17 delivery and 12 agriculture, with a concentration in counter-drone and search work. The United States holds 12 delivery and 4 defence documents. ### The operators with real delivery volume protect a different layer That absence reflects where the operators file. Zipline holds 61 intellectual-property assets spanning 2015 to 2025, covering fleet management, detect-and-avoid, docking and payload verification alongside route planning. Meituan files on trajectory planning and drone landing guidance among its applications. ## Areas - Delivery & Logistics · 1,135 documents - Defense & Security · 714 documents - Energy & Utilities · 592 documents - Public Safety & Disaster Response · 399 documents - Infrastructure & Construction · 319 documents - Agriculture · 179 documents