Beneath the forest canopy, trunks, branches, foliage, and shrubs are irregularly distributed, and passage widths constantly change. Lighting, wind disturbances, and occlusion also affect perception and control, placing greater demands on a drone’s ability to respond quickly.

Traditional flight control relies on manual modeling, parameter tuning, and fixed rules, which can lead to inflexible control strategies in forests. Training control policies with simulated and real-flight data offers a new approach to autonomous flight in complex environments.

By learning state features from perception data and generating control actions through machine learning or reinforcement learning, a drone may be able to keep flying, detect obstacles, and adjust its trajectory without relying on a precise map.

How can training data and reward mechanisms be designed to teach a drone stable control and real-time avoidance of trunks, branches, foliage, and shrubs in unfamiliar forests?

This is the focus of our featured innovation topic: drone forest flight based on reinforcement learning and machine learning, exploring the potential of learning-based flight control for safe autonomy in unstructured under-canopy environments.

 

01

Challenge Topic Overview

Drone Forest Flight Based on Reinforcement Learning and Machine Learning

This topic addresses forest environments with dense branches, tight spaces, and variable lighting. It explores how data-driven training, machine learning, and reinforcement learning can enable autonomous drone control and obstacle-avoiding flight.

Core tasks:

Collect flight, perception, and collision data from simulated and real forests to build reusable training datasets and environments;

Train flight control policies using supervised learning, imitation learning, or reinforcement learning to enable autonomous decisions on speed, attitude, and heading;

Perform online perception and safe avoidance of trunks, branches, foliage, and shrubs, and validate policy generalization in unfamiliar forests.

Required drone capabilities:

Multi-source sensor data acquisition and time synchronization;

Forest obstacle detection and spatial perception;

Data-driven flight control policy training;

Design of reinforcement learning environments, states, actions, and rewards;

Local path planning and real-time obstacle avoidance;

Sim-to-real transfer, flight safety, and performance evaluation.

The ultimate goal is to move from “manual parameter tuning and rule-based control” to “training autonomous flight policies with data and safely avoiding obstacles in forests.”

Challenge Objectives

Six technical areas form a complete loop from data collection and training to real-world forest flight validation.

Training Data and Environments

Give drones experience they can learn from. Build simulated and real-flight datasets containing viewpoints, depth, inertial measurements, control inputs, and collision labels.

Technical focus:

Data acquisition | Scene randomization | Training dataset construction

Perception and Feature Representation

Extract forest structure features suitable for control from camera, depth, or LiDAR data, identifying traversable space and nearby hazards.

Technical focus:

Multimodal fusion | Obstacle detection | Traversable area estimation

Learning-Based Flight Control

Use supervised learning or imitation learning to map perception to control, generating speed, attitude, or heading commands while maintaining stable flight.

Technical focus:

Behavior cloning | Policy networks | End-to-end control

Reinforcement Learning Policy Training

Design states, actions, and reward functions, and learn policies through repeated interaction that balance goal reaching, obstacle avoidance, safe clearance, and smooth control.

Technical focus:

Reward shaping | Policy optimization | Curriculum learning

Forest Obstacle Avoidance and Generalization

Adjust flight trajectories in real time across varying tree diameters, densities, foliage occlusion, and lighting conditions, and safely traverse forest areas not used for training.

Technical focus:

Local planning | Safety constraints | Domain randomization

Safety Validation and Evaluation

Test progressively in simulation, hardware-in-the-loop setups, and real flights. Record success rate, collision rate, minimum obstacle clearance, speed fluctuations, and inference latency.

Technical focus:

Sim-to-real transfer | Safety takeover | Metric-based evaluation

Application Scenarios

Learning-based autonomous forest flight is not only an intelligent control challenge; it also supports several real-world unmanned applications beneath the canopy.

Under-Canopy Inspection

Fly autonomously along forest passages to support inspections of forest facilities, trails, and key areas.

Forest Search and Rescue

Move rapidly through areas with tree occlusion and irregular paths, providing aerial support for target searches and guiding supply delivery.

Forestry Monitoring

Carry visible-light, infrared, or multispectral sensors to collect information on pests, diseases, and ecological conditions beneath the canopy.

Intelligent Flight Control Validation

Provide a comprehensive test scenario for learning-based control, sim-to-real transfer, and embodied intelligence algorithms.

Suggested validation outcomes: Using only onboard perception, the drone completes a specified flight route in a forest test area not used for training, continuously avoiding trunks, branches, foliage, and shrubs. Report flight trajectories, obstacle avoidance logs, collision rate, mission success rate, control smoothness, and other evaluation results.

02

Aerial Robotics Innovation and Creation Challenge

Advancing aerial robotics requires more validation in real-world scenarios. An innovative idea becomes a valuable product only through development, testing, and application validation.

To help more developers take part, the 2026–2027 Aerial Robotics Innovation and Creation Challenge is now officially open.

Open to universities, research institutes, technology companies, developer teams, and aerial robotics enthusiasts, the competition provides an innovation platform combining technical learning, project development, scenario-based validation, and project showcases.

Participating teams can develop projects using the FlyCore Integrated Aerial Robotics Control System, taking their own aerial robotics innovations from concept design and system integration through scenario-based validation and competitive project demonstrations.

The competition will also hold regular aerial robotics development bootcamps covering system integration, technical development, and hands-on projects, supporting learning, exchange, and project development for participating teams.

Ongoing Innovation Topics to Guide Exploration

To help teams move quickly into hands-on development, the challenge will continue to release a series of innovation topics.

Each topic will combine aerial robotics technologies with real-world applications, offering developers a reference for their projects.

Developers may research the recommended topics or extend them with innovations based on their own interests.

From a technical idea to a flying robot that can genuinely complete a task—this is the innovation process the challenge aims to promote.

 

03

Build with FlyCore and Aim for a Million in Innovation Funding

The FlyCore Integrated Aerial Robotics Control System provides an integrated development foundation spanning flight control, localization and navigation, and mission development.

FlyCore integrates core capabilities including flight control, mapping and localization, recognition and planning, and sensor connectivity. It forms a complete perception–localization–planning–control loop, helping developers rapidly build aerial robotics platforms, validate algorithms, and explore applications.

With FlyCore, participating teams can quickly complete concept design, feature development, and scenario-based validation, turning innovative ideas into robots that can fly and perform real tasks.

With FlyCore, teams can:

✓ Rapidly build aerial robotics platforms
✓ Validate algorithms and develop features
✓ Complete real-world testing
✓ Take part in the challenge’s competitive project demonstrations

Let ideas take flight, and bring innovation into real-world applications.