Sir Airffic Goບໍ່ຈຳເປັນຕ້ອງອ່ານທັງໝົດ. ກັ່ນຕອງຕາມຫົວຂໍ້ ແລ້ວອ່ານບົດສະຫຼຸບໜຶ່ງບົດແລະຄວາມໝາຍທາງທຸລະກິດຂອງມັນ.
ວິທີອ່ານບົດຄົ້ນຄວ້າ
- ບົດສະຫຼຸບ ຖືກຂຽນໃໝ່ໃຫ້ເດັກອາຍຸ 14 ປີອ່ານໄດ້. ກວດຕົ້ນສະບັບຖ້າຕ້ອງການຄວາມແມ່ນຍຳ.
- ຄວາມໝາຍທາງທຸລະກິດ ແມ່ນການຕີຄວາມຂອງທີມບັນນາທິການ.
- ຄວາມເລິກໃນການອ່ານ ບອກວ່າພວກເຮົາອ່ານເນື້ອຫາເຕັມຫຼືພຽງບົດຄັດຫຍໍ້.
ເບິ່ງບົດຄົ້ນຄວ້າທັງໝົດ (51)
Parallel Tracking and Mapping for Small AR Workspaces ↗
It showed how to split camera tracking (where am I?) and 3D map building into two programs running at the same time. That let a computer build a detailed map while still following the camera in real time.
ຄວາມໝາຍທາງຍຸດທະສາດ: Its tracking/mapping split became the template for later drone visual SLAM, underpinning GPS-free indoor flight products.
ຂໍ້ຈຳກັດ: It was designed for small desk-sized workspaces, not large outdoor areas or fast flight.
A Multi-State Constraint Kalman Filter for Vision-aided Inertial Navigation ↗
It proposed a filter (the MSCKF) that combines camera images with a motion sensor (IMU) to estimate a vehicle's position and orientation. It keeps accuracy high while keeping computation light.
ຄວາມໝາຍທາງຍຸດທະສາດ: As a lightweight visual-inertial navigation blueprint, it lives on in phone AR, drones and spacecraft localization.
ຂໍ້ຈຳກັດ: As a filter it slowly accumulates drift over long flights and needs separate corrections such as loop closure.
Minimum snap trajectory generation and control for quadrotors ↗
It showed how to compute smooth flight paths for a quadcopter through a list of waypoints in real time, avoiding jerky motion. Fast slalom flights in tight indoor spaces proved it works.
ຄວາມໝາຍທາງຍຸດທະສາດ: Minimum-snap trajectories remain a standard technique in many flight stacks, forming a base for agile autonomous flight.
ຂໍ້ຈຳກັດ: Experiments relied on external motion-capture for positioning rather than onboard sensing.
Trajectory generation and control for precise aggressive maneuvers with quadrotors ↗
It describes how to plan and control a quadcopter for precise, aggressive moves like flying through narrow gaps or perching. It is a landmark that showed the world how agile small drones can be.
ຄວາມໝາຍທາງຍຸດທະສາດ: By proving extreme agility, it raised expectations for businesses selling high-maneuverability drones for inspection, delivery and racing.
ຂໍ້ຈຳກັດ: It depends on precise external indoor tracking, so it does not transfer directly to real outdoor settings.
Towards A Swarm of Agile Micro Quadrotors ↗
The team built a 73-gram, 21-cm micro quadcopter and flew 20 of them together in formations. They showed that smaller drones are more agile and can fly in tight groups.
ຄວາມໝາຍທາງຍຸດທະສາດ: The idea of coordinating many tiny drones is one of the roots of today's drone light shows and swarm businesses.
ຂໍ້ຈຳກັດ: Positions came from external motion capture, so the drones did not sense their surroundings themselves.
Polynomial Trajectory Planning for Aggressive Quadrotor Flight in Dense Indoor Environments ↗
It showed a fast, numerically stable way to plan smooth polynomial flight paths for quadcopters in cluttered indoor spaces. The drone flew through tight indoor areas at up to 8 m/s.
ຄວາມໝາຍທາງຍຸດທະສາດ: Co-author Adam Bry later co-founded Skydio, making this a clear example of academic planning research flowing into commercial autonomous drones.
ຂໍ້ຈຳກັດ: It gives up guarantees of a globally optimal path and assumes a map is available.
Vision-Controlled Micro Flying Robots: From System Design to Autonomous Navigation and Mapping in GPS-Denied Environments ↗
It summarizes the European sFly project: small drones that fly on their own and build maps using only a camera and IMU, without GPS. It also covers sharing maps among several drones.
ຄວາມໝາຍທາງຍຸດທະສາດ: It showed camera-only small drones were nearing practicality, opening GPS-denied markets such as indoor and disaster sites.
ຂໍ້ຈຳກັດ: Computing limits of the time restricted flight speed and environmental complexity.
SVO: Fast Semi-Direct Monocular Visual Odometry ↗
SVO mixes two ways of reading images (raw pixel brightness and feature points) to estimate a single camera's motion very quickly. It was built to run on the weak computers inside small drones.
ຄວາມໝາຍທາງຍຸດທະສາດ: Fast, low-power pose estimation is critical on battery- and compute-limited drones, so SVO influenced many vehicle designs.
ຂໍ້ຈຳກັດ: A single camera cannot tell absolute scale, so other sensors such as an IMU are needed.
ORB-SLAM: a Versatile and Accurate Monocular SLAM System ↗
It is a single-camera system that tracks its own position and builds a map at the same time, from small rooms to large outdoor areas. It notices when it returns to a known place to fix errors, and its code is public.
ຄວາມໝາຍທາງຍຸດທະສາດ: As widely used open source, it became a shared base that lets drone and robot companies prototype visual SLAM quickly.
ຂໍ້ຈຳກັດ: The monocular version lacks absolute scale and can lose tracking in fast motion or texture-poor scenes.
Keyframe-based visual-inertial odometry using nonlinear optimization ↗
OKVIS fuses camera and IMU data by optimizing over a small set of important frames (keyframes) to estimate motion accurately. This tends to be more accurate than filter-based methods.
ຄວາມໝາຍທາງຍຸດທະສາດ: It moved optimization-based visual-inertial odometry toward practice, widening options for precise localization in industrial drones and AR.
ຂໍ້ຈຳກັດ: It is heavier to compute than filters, which strains small onboard computers.
Robust visual inertial odometry using a direct EKF-based approach ↗
ROVIO estimates motion by feeding the brightness of small image patches, together with IMU data, into an extended Kalman filter. It is lightweight and became popular on drones.
ຄວາມໝາຍທາງຍຸດທະສາດ: Lightweight, robust open-source estimators like this lowered development costs for research and industrial drones.
ຂໍ້ຈຳກັດ: It tracks a limited number of features and drifts over long distances.
Science, technology and the future of small autonomous drones ↗
This review looks ahead to small autonomous drones spreading into civilian work like farming, disaster response and delivery. It says the big hurdles are energy to stay in the air and the perception and decision-making needed in cluttered places.
ຄວາມໝາຍທາງຍຸດທະສາດ: By naming energy and perception/decision-making as the bottlenecks, it frames the core investment themes of aerial physical AI.
ຂໍ້ຈຳກັດ: It is a perspective piece and does not demonstrate new technology experimentally.
A Machine Learning Approach to Visual Perception of Forest Trails for Mobile Robots ↗
A deep neural network learned from photos of forest trails to tell which way the trail goes from a single image. It reached human-level accuracy and was used in early tests to fly a quadcopter along new trails.
ຄວາມໝາຍທາງຍຸດທະສາດ: It pioneered steering drones with perception learned from data instead of hand-written rules, a starting point for learning-based navigation.
ຂໍ້ຈຳກັດ: It only outputs rough directions and does not handle obstacle avoidance or altitude decisions.
A Benchmark and Simulator for UAV Tracking ↗
It released UAV123, 123 fully labeled HD videos shot from low-flying drones, plus a realistic drone simulator. This lets researchers compare which object trackers work best for drones.
ຄວາມໝາຍທາງຍຸດທະສາດ: A shared benchmark sped up comparison and improvement of video AI for follow-me filming and surveillance products.
ຂໍ້ຈຳກັດ: Its videos cover limited scene types, weather and times of day, not every real-world condition.
On-Manifold Preintegration for Real-Time Visual-Inertial Odometry ↗
It shows how to bundle many IMU readings between key images into one motion summary (preintegration), handling the math of rotations correctly. This makes accurate motion estimation possible in real time.
ຄວາມໝາຍທາງຍຸດທະສາດ: The method is built into many estimators such as VINS-Mono, making it a de facto industry building block.
ຂໍ້ຈຳກັດ: It assumes a correct IMU noise model, so accuracy drops when sensor behavior differs.
Toward Low-Flying Autonomous MAV Trail Navigation using Deep Neural Networks for Environmental Awareness ↗
NVIDIA researchers built a low-cost drone that flies low along forest trails. Its TrailNet neural network estimates the trail direction and how far the drone is from the center, and everything runs on an onboard Jetson TX1, enabling 1 km autonomous flights.
ຄວາມໝາຍທາງຍຸດທະສາດ: It showed drone perception could run fully on a small GPU edge board, giving chipmakers a foothold in the drone market.
ຂໍ້ຈຳກັດ: It targets relatively simple trail settings, not complex urban spaces or high-speed flight.
CAD2RL: Real Single-Image Flight Without a Single Real Image ↗
A drone-flying AI was trained with reinforcement learning only in 3D CAD-based simulation, then flew a real quadcopter indoors without hitting things, never seeing a real photo in training. The trick is to randomly vary how the simulated rooms look.
ຄວາມໝາຍທາງຍຸດທະສາດ: It established the train-in-sim, deploy-in-reality (domain randomization) path that cuts the cost of collecting real flight data.
ຂໍ້ຈຳກັດ: It covers fairly slow indoor flight, with no proof for high-speed or outdoor use.
AirSim: High-Fidelity Visual and Physical Simulation for Autonomous Vehicles ↗
Microsoft built AirSim, a simulator for drones and cars that uses a game engine (Unreal Engine) to look and behave realistically. They checked it against real quadcopter flights.
ຄວາມໝາຍທາງຍຸດທະສາດ: A high-quality open-source simulator made AI training data and safe testing cheaper, lowering barriers to drone development.
ຂໍ້ຈຳກັດ: A gap between simulation and reality remains, so real-world testing is still required.
Reluplex: An Efficient SMT Solver for Verifying Deep Neural Networks ↗
It proposed Reluplex, a way to mathematically prove that a neural network will always behave a certain way. It was tested on a neural-network version of ACAS Xu, a collision-avoidance system for unmanned aircraft, handling networks ten times larger than before.
ຄວາມໝາຍທາງຍຸດທະສາດ: Certifying AI for aviation requires proof, so this points the way to getting drone collision-avoidance AI accepted by regulators.
ຂໍ້ຈຳກັດ: Network size is still limited, so it is hard to apply directly to large camera-based perception AI.
Crazyswarm: A Large Nano-Quadcopter Swarm ↗
Crazyswarm is a hardware and software system for flying dozens of palm-sized Crazyflie drones together. It lets researchers try large swarm flights easily.
ຄວາມໝາຍທາງຍຸດທະສາດ: A cheap, open swarm testbed sped up swarm-algorithm research and talent development worldwide.
ຂໍ້ຈຳກັດ: It mainly relies on indoor setups like motion capture, unlike real outdoor swarm operations.
VINS-Mono: A Robust and Versatile Monocular Visual-Inertial State Estimator ↗
VINS-Mono estimates a vehicle's position and orientation, at real-world scale, using just one camera and a cheap IMU. It handles startup, failure recovery and loop detection, was tested on real drone flights and an iPhone, and is open source.
ຄວາມໝາຍທາງຍຸດທະສາດ: As open-source, high-accuracy localization from cheap sensors, it became a standard component for many drone companies and labs.
ຂໍ້ຈຳກັດ: Accuracy drops in darkness, on textureless walls or under sudden lighting changes.
DroNet: Learning to Fly by Driving ↗
DroNet, a light 8-layer neural network, looks at one camera image and outputs a steering direction and a chance of collision so a drone can fly through streets. It learned from videos taken by cars and bicycles, avoiding risky drone data collection.
ຄວາມໝາຍທາງຍຸດທະສາດ: It showed drone AI can be built from other domains' data (driving), cutting data-collection costs.
ຂໍ້ຈຳກັດ: It is reactive steering without maps or planning, so goal-directed missions need extra systems.
The Unmanned Aerial Vehicle Benchmark: Object Detection and Tracking ↗
UAVDT takes about 80,000 frames from roughly 10 hours of drone video and labels vehicles with boxes plus attributes like weather and altitude. It showed top methods struggle with tiny, crowded objects and moving cameras.
ຄວາມໝາຍທາງຍຸດທະສາດ: For aerial video analytics like traffic monitoring, ground-view AI does not transfer directly, so drone-specific data is a competitive asset.
ຂໍ້ຈຳກັດ: It mainly covers vehicles, with limited other object types.
DOTA: A Large-Scale Dataset for Object Detection in Aerial Images ↗
DOTA is a large dataset for finding planes, ships, cars and more in overhead images. Objects are labeled with rotated boxes so tilted objects can be marked precisely.
ຄວາມໝາຍທາງຍຸດທະສາດ: It became a standard benchmark for overhead detection, the basis for accuracy competition in satellite and drone image analytics.
ຂໍ້ຈຳກັດ: It is mostly aerial/satellite imagery, which looks different from low-altitude oblique drone video.
Optimized flocking of autonomous drones in confined environments ↗
It made dozens of drones fly together outdoors like a flock of birds, with no central commander. An evolutionary optimizer tuned the movement rules so the flock could move through confined spaces without crashing.
ຄວາມໝາຍທາງຍຸດທະສາດ: Proving decentralized flocking outdoors made multi-drone operations for inspection, search and farming more realistic.
ຂໍ້ຈຳກັດ: It relies on GPS and radio links for sharing positions, so it is weak where GPS or communication fails.
Radar micro-Doppler signatures of drones and birds at K-band and W-band ↗
To tell drones from birds by radar, it measured and compared the tiny speed wobbles (micro-Doppler) made by propellers and flapping wings in two radar frequency bands. Spinning rotors show features that birds do not.
ຄວາມໝາຍທາງຍຸດທະສາດ: Bird false alarms are a top problem for counter-drone radar, so classification is a differentiator for airport and critical-site products.
ຂໍ້ຈຳກັດ: Measurements cover limited drone types and conditions, so long-range or crowded-sky performance needs further checks.
MAVBench: Micro Aerial Vehicle Benchmarking ↗
MAVBench is a simulator plus benchmark apps to study how computing speed, power use and battery life interact in autonomous small drones. It showed smarter computer design can improve battery use by up to 1.8x.
ຄວາມໝາຍທາງຍຸດທະສາດ: It showed faster compute can save total energy by finishing missions sooner, supporting the business case for drone-specific chips.
ຂໍ້ຈຳກັດ: Much of the evaluation is in simulation, which differs from varied real-world conditions.
A 64mW DNN-based Visual Navigation Engine for Autonomous Nano-Drones ↗
It put an ultra-low-power chip on a palm-sized nano-drone and ran the DroNet neural network onboard using only 64 milliwatts, enabling autonomous flight.
ຄວາມໝາຍທາງຍຸດທະສາດ: It made AI autonomy possible on tens-of-grams drones, pointing to a new market for ultra-low-power edge AI chips.
ຂໍ້ຈຳກັດ: Only small networks fit, which strongly limits decision-making and speed in complex environments.
Machine Learning-Based Drone Detection and Classification: State-of-the-Art in Research ↗
This review organizes machine-learning research on detecting and classifying drones by four sensing methods: radar, sound, cameras and radio signals (RF). It compares the strengths and weaknesses of each.
ຄວາມໝາຍທາງຍຸດທະສາດ: By laying out each sensor's limits, it supports the business need for multi-sensor counter-drone systems.
ຂໍ້ຈຳກັດ: As a 2019 review, it misses the rapid deep-learning progress that followed.
RF-based drone detection and identification using deep learning approaches: An initiative towards a large open source drone database ↗
It built an open database of radio (RF) signals between drones and their controllers, and used deep learning to detect drones and identify their type and flight mode.
ຄວາມໝາຍທາງຍຸດທະສາດ: RF detection is cheap and passive, so open data made it easier to develop RF-based counter-drone products.
ຂໍ້ຈຳກັດ: It struggles with drones that fly autonomously without radio links or models not in the database.
Event-based Vision: A Survey ↗
This review explains event cameras, which report only pixels whose brightness changes, instantly, instead of taking whole frames at fixed times. It covers how they work, algorithms and uses, and why they handle very fast motion and extreme lighting.
ຄວາມໝາຍທາງຍຸດທະສາດ: It laid out the basics of a sensor suited to fast flight and harsh lighting, opening the door to event-camera drone products.
ຂໍ້ຈຳກັດ: Event cameras are still costly and lower-resolution, and ordinary camera AI cannot be reused directly.
Dynamic obstacle avoidance for quadrotors with event cameras ↗
Using event cameras that report only brightness changes, a quadrotor dodges fast objects such as a thrown ball. The reaction delay is far shorter than with ordinary cameras.
ຄວາມໝາຍທາງຍຸດທະສາດ: It showed low-latency sensors can become a safety differentiator where moving obstacles are common.
ຂໍ້ຈຳກັດ: It was shown in limited conditions; cluttered outdoor scenes with many objects remain under-tested.
Deep Drone Acrobatics ↗
A neural network trained only in simulation flies a real quadrotor through acrobatic maneuvers such as loops and flips. It maps camera and inertial data directly to control.
ຄວາມໝາຍທາງຍຸດທະສາດ: It showed sim-to-real training can transfer advanced flight skills, lowering the cost of teaching drones.
ຂໍ້ຈຳກັດ: It performs set maneuvers in set settings; generalization to unknown places or long missions is not shown.
Flightmare: A Flexible Quadrotor Simulator ↗
Flightmare is a fast, flexible drone simulator that separates physics from game-engine rendering. It can run many drones in parallel to speed up reinforcement learning.
ຄວາມໝາຍທາງຍຸດທະສາດ: Open, fast simulators widen access to autonomy development and cut the cost of training data.
ຂໍ້ຈຳກັດ: Gaps in visuals and aerodynamics versus reality remain, so real-flight testing is still required.
Radar-Inertial Ego-Velocity Estimation for Visually Degraded Environments ↗
It combines a small radar with inertial sensing to estimate a drone's own velocity when smoke, dust or darkness blinds cameras.
ຄວາມໝາຍທາງຍຸດທະສາດ: It shows radar as a navigation sensor, pointing to demand for drones and radar parts in underground and disaster sites.
ຂໍ້ຈຳກັດ: It estimates velocity, but position drift accumulates, so long flights need other aids.
Evaluating collision avoidance for small UAS using ACAS X ↗
It simulates many close encounters between small drones and crewed aircraft to compare who should avoid whom. It introduces ACAS X-based avoidance rules usable by small drones with weak climb ability.
ຄວາມໝາຍທາງຍຸດທະສາດ: It shows standardized avoidance logic (ACAS sXu) is a prerequisite for certifying beyond-visual-line-of-sight flight.
ຂໍ້ຈຳກັດ: It is simulation-based, simplifying real sensor errors and communication delays.
Learning high-speed flight in the wild ↗
An AI that learned in simulation by imitating an expert planner flies through forests and around buildings at up to about 10 m/s, avoiding obstacles. It decides directly from sensors without building a map.
ຄວາມໝາຍທາງຍຸດທະສາດ: Map-free learned flight hints at faster, cheaper drones for inspection and search.
ຂໍ້ຈຳກັດ: Success drops with speed and clutter, and guaranteeing safety on failure is unsolved.
Time-optimal planning for quadrotor waypoint flight ↗
The method computes the fastest path through a series of waypoints using the drone's full performance limits. In tests it flew faster than expert human pilots.
ຄວາມໝາຍທາງຍຸດທະສາດ: Planning at the vehicle's limits can shorten delivery and inspection times and save battery.
ຂໍ້ຈຳກັດ: It is computationally heavy and unsuited to on-the-fly replanning or unknown obstacles.
EGO-Planner: An ESDF-free Gradient-based Local Planner for Quadrotors ↗
This light planner skips computing a full distance map and fixes the path only where it would collide. It runs fast even on small onboard computers.
ຄວາມໝາຍທາງຍຸດທະສາດ: Released as open code, it became a widely used basis for autonomous avoidance on low-cost drones.
ຂໍ້ຈຳກັດ: Being local, it can detour or fail in large maze-like spaces with dead ends.
EGO-Swarm: A Fully Autonomous and Decentralized Quadrotor Swarm System in Cluttered Environments ↗
Each drone plans for itself and avoids the others, forming a swarm with no central controller. Several drones flew through forest-like clutter.
ຄວາມໝາຍທາງຍຸດທະສາດ: Decentralized swarming brings search and inspection in weak-communication areas closer to practice.
ຂໍ້ຈຳກັດ: Only a few drones were used; large swarms and jammed links are untested.
Detection and Tracking Meet Drones Challenge ↗
The paper presents VisDrone, a large set of drone images and videos labeled with people, cars and more, and results from its challenges. It shows how hard it is to find small objects seen from above.
ຄວາມໝາຍທາງຍຸດທະສາດ: It became a common yardstick for aerial vision AI, spurring competition among detection vendors.
ຂໍ້ຈຳກັດ: Data come mostly from Chinese cities, so results may not carry to other regions, weather or night.
Real-Time Drone Detection and Tracking With Visible, Thermal and Acoustic Sensors ↗
The system combines visible and thermal cameras with microphones to detect and track drones. It also released data useful for telling drones from birds and planes.
ຄວາມໝາຍທາງຍຸດທະສາດ: It supports multi-sensor fusion as the basic form of counter-drone products, and its open data aid detector development.
ຂໍ້ຈຳກັດ: Detection range is a few hundred meters, short of military ranges or swarm detection.
Swarm of micro flying robots in the wild ↗
A swarm of palm-sized drones flies through dense forest such as bamboo groves without GPS or external positioning. Each drone decides using its own camera and computer.
ຄວາມໝາຍທາງຍຸດທະສາດ: It showed small, cheap drones can fly autonomously as a swarm, boosting civil and defense swarm uses.
ຂໍ້ຈຳກັດ: Short flight time and tiny payload keep it short of real missions.
FAST-LIO2: Fast Direct LiDAR-inertial Odometry ↗
It combines LiDAR point clouds with inertial sensing to estimate position and map the surroundings quickly and accurately. It uses raw points directly instead of extracting features.
ຄວາມໝາຍທາງຍຸດທະສາດ: Widely adopted as open source, it underpins LiDAR drone inspection where GPS is unavailable.
ຂໍ້ຈຳກັດ: LiDAR is heavy and costly, and drift grows in featureless places such as long corridors.
FASTER: Fast and Safe Trajectory Planner for Navigation in Unknown Environments ↗
In unknown places it plans a fast path using both seen-free and unseen space, while always keeping a backup path inside known-safe space. It balances speed and safety.
ຄວາມໝາຍທາງຍຸດທະສາດ: It set the fast-but-guaranteed-safe planning pattern later used by high-speed systems such as SUPER.
ຂໍ້ຈຳກັດ: Its guarantees assume correct localization and maps; sensor errors break them.
Champion-level drone racing using deep reinforcement learning ↗
Swift, an AI trained with reinforcement learning, beat world-class human drone-racing pilots using only onboard cameras and computing. It is a case of AI beating human champions in a physical sport.
ຄວາມໝາຍທາງຍຸດທະສາດ: It showed learned autonomy can surpass human piloting, raising expectations for AI flight in defense and fast inspection.
ຂໍ້ຈຳກັດ: It was trained for a fixed track and lighting, and performance drops when conditions change.
AerialVLN: Vision-and-Language Navigation for UAVs ↗
It created a task and dataset where a drone flies through a city following spoken-style instructions such as going to a building roof. Existing AI did far worse than humans.
ຄວາມໝາຍທາງຍຸດທະສາດ: It set a benchmark for language-directed drones, a starting point for bringing language AI to the air.
ຂໍ້ຈຳກັດ: Evaluation is in simulation; real drones in real cities were not tested.
RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control ↗
A large AI trained on web images and text is taught to output robot actions as if they were words. It could handle some objects and instructions it was never taught.
ຄວາມໝາຍທາງຍຸດທະສາດ: It set the template for vision-language-action foundation models that drone AI efforts now follow.
ຂໍ້ຈຳກັດ: It was tested on ground robot arms and runs too slowly for fast flight control as is.
ChatGPT for Robotics: Design Principles and Model Abilities ↗
It lays out how to have ChatGPT turn everyday-language instructions into code that controls robots, including drones. Examples include drone inspection in a simulator.
ຄວາມໝາຍທາງຍຸດທະສາດ: It suggested non-experts could set up drone tasks in plain language, changing how control software competes.
ຂໍ້ຈຳກັດ: It assumes a human checks outputs; no mechanism safely catches language-model errors.
OpenFly: A Versatile Toolchain and Large-scale Benchmark for Aerial Vision-Language Navigation ↗
OpenFly releases a toolchain that automatically generates large amounts of data from many virtual cities and scenes, plus a big benchmark for language-guided drone navigation.
ຄວາມໝາຍທາງຍຸດທະສາດ: Automating data generation targets the biggest barrier for drone foundation models: lack of training data.
ຂໍ້ຈຳກັດ: Its data are mostly synthetic, so real-flight performance is uncertain.
Safety-assured high-speed navigation for MAVs ↗
SUPER, a small LiDAR drone, flies safely through forests at roughly 70 km/h without GPS or light. It always keeps a safe backup path while choosing a fast one.
ຄວາມໝາຍທາງຍຸດທະສາດ: Combining speed with safety brings dark, fast missions like night search and inspection closer to use.
ຂໍ້ຈຳກັດ: LiDAR adds weight and cost, and very thin obstacles such as wires remain hard.
ບໍ່ພົບບັນທຶກທີ່ກົງກັນ. ລອງໃຊ້ຄຳອື່ນ.