Source-linked AI summary
Fast, Autonomous Flight in GPS-Denied and Cluttered Environments
Kartik Mohta, Michael Watterson, Yash Mulgaonkar, Sikang Liu, Chao Qu, Anurag Makineni, Kelsey Saulnier, Ke Sun, Alex Zhu, Jeffrey Delmerico, Konstantinos Karydis, Nikolay Atanasov, Giuseppe Loianno, Davide Scaramuzza, Kostas Daniilidis, Camillo Jose Taylor, Vijay Kumar
TL;DR
The paper addresses fast, reliable quadrotor navigation between target poses in cluttered, GPS-denied environments without ground-truth infrastructure. It presents an integrated onboard system for sensing, estimation, control, mapping, planning, and trajectory generation, and demonstrates successful autonomous operation in laboratory and warehouse obstacle courses. The system completed varied obstacle courses with only a relative goal input, including runs where the monitoring communication link was lost.
Problem
Reliable high-speed navigation in dynamic, partially known, GPS-denied environments requires fully autonomous systems based only on onboard sensing and computation.
Method
The paper integrates onboard state estimation, control, mapping, search-based planning, and constrained trajectory generation for quadrotor navigation.
Results
The system successfully navigated varied obstacle courses in laboratory and DARPA FLA warehouse environments using only the goal position relative to the start.
Takeaways & Limitations
The integrated system supports autonomous quadrotor operation in unknown, cluttered environments without human interaction after launch.
Takeaways & Limitations
Performance is constrained by visual-estimation drift during 90° aisle transitions, a limited stereo baseline, and local-map planning time that limits speed.
Abstract
from arXiv · showhide
One of the most challenging tasks for a flying robot is to autonomously navigate between target locations quickly and reliably while avoiding obstacles in its path, and with little to no a-priori knowledge of the operating environment. This challenge is addressed in the present paper. We describe the system design and software architecture of our proposed solution, and showcase how all the distinct components can be integrated to enable smooth robot operation. We provide critical insight on hardware and software component selection and development, and present results from extensive experimental testing in real-world warehouse environments. Experimental testing reveals that our proposed solution can deliver fast and robust aerial robot autonomous navigation in cluttered, GPS-denied environments.
1 Introduction
The paper targets fast, autonomous quadrotor navigation between poses in cluttered, GPS-denied environments using only onboard sensing and computation. It integrates estimation, control, mapping, planning, and trajectory generation, with experiments in obstacle-rich environments.
- Real-world navigation requires onboard sensing because dynamic, partially known, and GPS-denied environments lack reliable ground-truth measurements.
- The system is designed to navigate at high speeds between start and goal poses while estimating state, controlling motion, mapping, and planning onboard.
- The navigation pipeline uses visual odometry, sensor fusion, local mapping, search-based planning, and trajectory generation with dynamic constraints.
- Trajectory generation uses a quadratic program constrained by safe corridors and limits on velocity, acceleration, and jerk.
- The complete system was tested in laboratory and DARPA FLA obstacle-rich environments to evaluate autonomous navigation without a known map.
2 System Design
The system design combines a lightweight high-performance quadrotor platform with stereo vision, nodding-lidar mapping, onboard computation, wireless communication, and ROS-based software modules. Component choices balance sensing, computation, mass, power, and operational robustness.
- Platform Design: The platform design prioritizes fast, agile flight while satisfying acceleration, thrust-to-weight, sensing, and payload constraints.
- Platform Design: The selected platform uses a DJI Flamewheel 450 frame, E600 propulsion, and Pixhawk autopilot, weighing 2.1 kg with sensing and computation payload but without the battery.
- Sensing, Computation and Communication: A 2D Hokuyo lidar mounted on a one-degree-of-freedom nodding gimbal provides the 3D map required for planning while limiting weight.
- Sensing, Computation and Communication: An Intel NUC i7 runs the full onboard estimation, control, mapping, and planning software stack while supporting sensor-data recording.
- Software Architecture: ROS organizes estimation, control, mapping, and planning components and allows nodes to be distributed between the robot and workstation for development.
3 Estimation and Control
The system uses SVO for visual motion estimation and combines sparse photometric alignment, feature correspondence, pose refinement, and depth mapping. Its estimation design balances robustness and computational efficiency through stereo sensing, recursive depth filtering, and local optimization.
- Sensor choice: Stereo visual odometry is selected because it improves robustness over monocular sensing while remaining simpler than multi-camera systems.Stereo can also fall back to a single camera when matching is impossible for distant features.
- Motion estimation: SVO combines feature-based and direct methods to estimate six-degree-of-freedom camera motion from salient image features and photometric alignment.It aligns images using sparse features, obtains subpixel correspondences, and refines pose using reprojection error.
- Motion estimation: Feature alignment minimizes patch intensity differences against a reference patch, using two-dimensional corrections for corners and edge-normal corrections for edge features.The alignment step can introduce reprojection error, typically below 0.5 pixels, which is corrected during pose and landmark refinement.
- Motion estimation: Local optimization of the latest camera pose and 3D points is used instead of whole-trajectory bundle adjustment to permit more efficient MAV motion estimation.Whole-trajectory optimization provides higher accuracy, but the local strategy is considered sufficient for this application.
- Mapping: The mapping thread estimates feature depth with recursive Bayesian filters, searches along epipolar lines, and initializes map points once depth uncertainty falls below a threshold.Depth filters are initialized at corner and edge features in newly selected keyframes and later support motion estimation.
- State estimation: The state representation includes world-frame position and velocity, roll, pitch, yaw, accelerometer bias, and gyroscope bias, using ZYX Euler angles.The paper assumes expected roll and pitch remain below 90°, avoiding gimbal lock in this application.
4 Mapping and Planning
The system combines local and global mapping with replanning, safe-corridor path modification, and polynomial trajectory optimization to navigate efficiently around obstacles. Its planner addresses dead ends while the trajectory generator balances path adherence, dynamic feasibility, and obstacle clearance.
- Framework: The navigation framework sends a goal to the planner, converts the resulting path into a trajectory, and derives desired states for the robot controller at 200 Hz.
- Mapping: The mapping system uses a fine-resolution local 3D occupancy map for recent measurements and a coarse global 2D information map to address dead ends.The local map limits accumulated mapping error, while the global map records explored space and detected walls.
- Planning: A⋆ plans paths on a hybrid graph linking local 3D voxels with the global information map, producing paths that remain globally consistent.The method is illustrated on a dead-end corridor where newly observed geometry changes the planned route.
- Trajectory generation: Trajectory generation inflates free space around the path into a safe corridor, shifts intermediate waypoints toward its center, and produces a dynamically feasible trajectory.Regional inflation excludes occupied map points before trajectory optimization.
- Trajectory generation: The trajectory optimizer adds a distance-to-path cost weighted by ϵ, allowing trajectory shape to remain close to the modified path while avoiding obstacles.Increasing ϵ makes the generated trajectory closer to the given path.
- Trajectory generation: The implementation represents trajectories as nth-order polynomial splines, uses n = 7 in experiments, and enforces continuity through 3 derivatives.The trajectory is sampled for inequality constraints and centering costs, with segment times based on fitted trapezoidal velocity profiles.
5 Experimental Results
Experiments evaluated the fully autonomous system in laboratory and warehouse-like obstacle courses, including slalom, aisle transitions, and high-speed flight. The system succeeded in several challenging tasks, while the 90° transition exposed vision-related drift and the high-speed run reached 7 m/s over 65 m.
- State estimation: The estimator was tested against motion-capture ground truth during aggressive manual flight reaching 4 m/s and accelerations of 4 m/s2.
- State estimation: SVO ran onboard with forward-facing stereo cameras and maintained feature tracking and metrically consistent sparse mapping during high-speed warehouse maneuvers.A separate straight-line demonstration reached over 15 m/s across 50 m.
- Obstacle-course experiments: The complete system operated fully autonomously in real-world warehouse-like obstacle courses without human interaction after takeoff.The courses were approximately 65 m long, with 3 m-wide aisles and scaffolding, towers, boxes, and scissor lifts.
- Slalom: In the slalom task, the projected robot map matched the actual obstacle-course layout while the robot followed the required zigzag route.
- Aisle transitions: In the 45° aisle transition, the robot successfully moved into the second aisle, reached the goal, and produced a projected map matching the actual course layout.
- Aisle transitions: The 90° aisle transition reached the goal but performed poorly because feature loss caused vision jumps, yaw drift, and resulting position errors.The problem arose when sideways motion left the system with distant, poorly triangulated features; forward motion restored better estimates.
- High-speed flight: 7 m/s was achieved over a 65 m obstacle-free flight to the desired goal, with estimated net position drift below 2 m and adequate aggressive-trajectory tracking.
6 Discussion and Conclusion
The discussion identifies modular design, sensor selection, and local-map planning as important to autonomous quadrotor navigation, while warehouse tests expose limits from visual-odometry drift and planning-map size. The system achieved onboard autonomous navigation in cluttered environments, but substantially higher speeds require further work across its modules.
- Design lessons: The modular architecture enabled parallel development and testing of planner, estimation, and control components.ROS and message-passing interfaces allowed planner development in simulation while estimation and control were developed separately.
- Experimental results: The full system reached speeds of up to 7 m/s during control testing in an empty aisle.Figure 24 reports control performance while running the full navigation system.
- Design lessons: Stereo sensing was selected to improve visual-odometry robustness compared with monocular sensing.The paper notes that sensor configuration affects robustness and that stereo provides increased robustness over a monocular setup.
- Design lessons: Local-map planning tolerates state-estimation drift because drift remains small during the short period in which the local map is constructed.Maintaining a global map would require explicit global consistency and loop closures to eliminate drift.
- Limitations and future work: Visual-odometry drift and local-map planning time constrained performance, while reaching the 20 m/s target requires improvements in estimation, control, mapping, and planning.The paper attributes the speed limit to poor depth estimates, bounded map size, and the computational cost of A* planning.
- Experimental results: Warehouse and laboratory tests showed successful navigation through obstacle courses, including runs completed after the base-station communication link was lost.The operator supplied only the goal position relative to the starting position, while the base station was used for monitoring.