MULTI-SENSOR FUSION FOR ROBUST AUTONOMOUS FLIGHT IN INDOOR AND OUTDOOR ENVIRONMENTS WITH A ROTORCRAFT MICRO-AERIAL VEHICLE (MAV)
The subject matter described herein includes a modular and extensible approach to integrate noisy measurements from multiple heterogeneous sensors that yield either absolute or relative observations at different and varying time intervals, and to provide smooth and globally consistent estimates of position in real time for autonomous flight. We describe the development of the algorithms and software architecture for a new 1.9 kg MAV platform equipped with an IMU, laser scanner, stereo cameras, pressure altimeter, magnetometer, and a GPS receiver, in which the state estimation and control are performed onboard on an Intel NUC 3rd generation i3 processor. We illustrate the robustness of our framework in large-scale, indoor-outdoor autonomous aerial navigation experiments involving traversals of over 440 meters at average speeds of 1.5 m/s with winds around 10 mph while entering and exiting buildings.
This application claims the benefit of U.S. Provisional Application Ser. No. 61/910,022, filed Nov. 27, 2013, the disclosure of which is incorporated herein by reference in its entirety.
GOVERNMENT INTERESTThis invention was made with government support under Grant Nos. N00014-07-1-0829, N00014-08-1-0696, N00014-09-1-1031, and N00014-09-1-1051 awarded by the Office of Naval Research, Grant Nos. 1138847, 113830, and 1138110 awarded by the National Science Foundation, Grant Nos. W911NF-08-2-0004 and W911NF-10-2-0016 awarded by the Air Force Office of Scientific Research, and Grant No. FA9550-10-1-0567 awarded by the Army Research Laboratory. The government has certain rights in the invention.
TECHNICAL FIELDThe subject matter described herein relates to controlling autonomous flight in a micro-aerial vehicle. More particularly, the subject matter described herein relates to multi-sensor fusion for robust autonomous flight in indoor and outdoor environments with a rotorcraft micro-aerial vehicle (MAV).
BACKGROUNDMicro-aerial vehicles, such as rotorcraft micro-aerial vehicles, are capable of flying autonomously. Accurate autonomous flight can be achieved provided that there is sufficient sensor data available to provide control input for the autonomous flight. For example, in some outdoor environments where a global positioning system (GPS) is available, autonomous flight can be achieved based on GPS signals. However, in environments where GPS is not available, such as indoor environments and even outdoor urban environments, autonomous flight based on GPS alone is not possible. In some indoor environments, magnetometer output may not be available or reliable due to magnetic interference caused by structures. Thus, reliance on a single modality of sensor to control flight of a rotorcraft MAV may not be desirable.
Another goal of controlling autonomous flight of a rotorcraft MAV is smooth transition between states when a sensor modality that was not previously available becomes available. For example, when a rotorcraft MAV is flying indoors where GPS is not available and then transitions to an outdoor environment where GPS suddenly becomes available, the rotorcraft may determine that it is far off course and may attempt to correct the error by immediately moving to be on course. It is desirable that such transitions be smooth, rather than having the rotorcraft immediately make large changes in velocity and trajectory to get back on course.
Multiple types of sensor data are available to control autonomous flight in rotorcraft micro-aerial vehicles. For example, onboard cameras, laser scanners, GPS transceivers, and accelerometers can provide multiple inputs that are suitable as control inputs for controlling flight. However, as stated above, relying on any one of these sensors fails when the assumptions associated with the sensor fails. Because each type of sensor produces a unique kind of output with a unique level of uncertainty in its measurement, there exists a need for improved methods, systems, and computer readable media for multi-sensor fusion for robust autonomous flight in indoor and outdoor environments with a rotorcraft MAV.
SUMMARYThe subject matter described herein includes a modular and extensible approach to integrate noisy measurements from multiple heterogeneous sensors that yield either absolute or relative observations at different and varying time intervals, and to provide smooth and globally consistent estimates of position in real time for autonomous flight. We describe the development of the algorithms and software architecture for a new 1.9 kg MAV platform equipped with an inertial measurement unit (IMU), laser scanner, stereo cameras, pressure altimeter, magnetometer, and a GPS receiver, in which the state estimation and control are performed onboard on an Intel NUC 3rd generation i3 processor. We illustrate the robustness of our framework in large-scale, indoor-outdoor autonomous aerial navigation experiments involving traversals of over 440 meters at average speeds of 1.5 m/s with winds around 10 mph while entering and exiting buildings.
The subject matter described herein may be implemented in hardware, software, firmware, or any combination thereof. As such, the terms “function”, “node” or “module” as used herein refer to hardware, which may also include software and/or firmware components, for implementing the feature being described. In one exemplary implementation, the subject matter described herein may be implemented using a computer readable medium having stored thereon computer executable instructions that when executed by the processor of a computer control the computer to perform steps. Exemplary computer readable media suitable for implementing the subject matter described herein include non-transitory computer-readable media, such as disk memory devices, chip memory devices, programmable logic devices, and application specific integrated circuits. In addition, a computer readable medium that implements the subject matter described herein may be located on a single device or computing platform or may be distributed across multiple devices or computing platforms.
The subject matter described herein will now be explained with reference to the accompanying drawings of which:
Rotorcraft micro-aerial vehicles (MAVs) are ideal platforms for surveillance and search and rescue in confined indoor and outdoor environments due to their small size, superior mobility, and hover capability. In such missions, it is essential that the MAV is capable of autonomous flight to minimize operator workload. Robust state estimation is critical to autonomous flight especially because of the inherently fast dynamics of MAVs. Due to cost and payload constraints, most MAVs are equipped with low cost proprioceptive sensors (e.g. MEMS IMUs) that are incapable for long term state estimation. As such, exteroceptive sensors, such as GPS, cameras, and laser scanners, are usually fused with proprioceptive sensors to improve estimation accuracy. Besides the well-developed GPS-based navigation technology [1, 2]. There is recent literature on robust state estimation for autonomous flight in GPS-denied environments using laser scanners [3, 4], monocular camera [5, 6], stereo cameras [7, 81, and RGB-D sensors [9]. However, all these approaches rely on a single exteroceptive sensing modality that is only functional under certain environment conditions. For example, laser-based approaches require structured environments, vision based approaches demand sufficient lighting and features, and GPS only works outdoors. This makes them prone to failure in large-scale environments involving indoor-outdoor transitions, in which the environment can change significantly. It is clear that in such scenarios, multiple measurements from GPS, cameras, and lasers may be available, and the fusion of all these measurements yields increased estimator accuracy and robustness. In practice, however, this extra information is either ignored or used to switch between sensor suites [10].
The main goal of this work is to develop a modular and extensible approach to integrate noisy measurements from multiple heterogeneous sensors that yield either absolute or relative observations at different and varying time intervals, and to provide smooth and globally consistent estimates of position in real time for autonomous flight. The first key contribution, that is central to our work, is a principled approach, building on [11], to fusing relative measurements by augmenting the vehicle state with copies of previous states to create an augmented state vector for which consistent estimates are obtained and maintained using a filtering framework. A second significant contribution is our Unscented Kalman Filter (UKF) formulation in which the propagation and update steps circumvent the difficulties that result from the semi-definiteness of the covariance matrix for the augmented state. Finally, we demonstrate results with a new experimental platform (
Next, we present previous work on which our work is based. In Section III we outline the modeling framework before presenting the key contributions of UKF-based sensor fusion scheme in Section IV. We bring all the ideas together in our description of the experimental platform and the experimental results in Section VI.
II. Previous WorkWe are interested in applying constant computation complexity filtering-based approaches, such as nonlinear variants of the Kalman filter, to fuse all available sensor information. We stress that although SLAM-based multi-sensor fusion approaches [12, 13] yield optimal results, they are computationally expensive for real-time state feedback for the purpose of autonomous control.
While it is straightforward to fuse multiple absolute measurements such as GPS, pressure/laser altimeter in a recursive filtering formulation, the fusion of multiple relative measurements obtained from laser or visual odometry are more involved. It is common to accumulate the relative measurements with the previous state estimates fuse them as pseudo-absolute measurements [5, 14]. However, such fusion is sub-optimal since the resulting global position and yaw covariance is inconsistently small compared to the actual estimation error. This violates the observability properties [6], which suggests that such global quantities are in fact unobservable. As such, we develop our method based on state augmentation techniques [11] to properly account for the state uncertainty when applying multiple relative measurements from multiple sensors.
We aim to develop a modular framework that allows easy addition and removal of sensors with minimum coding and mathematical derivation. We note that in the popular EKF-based formulation [5, 8], the computation of Jacobians can be problematic for complex systems like MAVs. As such, we employ a loosely coupled, derivative-free Unscented Kalman Filter (UKF) framework [1]. Switching from EKF to UKF poses several challenges, which will be detailed and addressed in Sect. IV-A. [15] is similar to our work. However, the EKF-based estimator in [15] does not support fusion of multiple relative measurements.
III. Multi-Sensor System ModelWe define vectors in the world and body frames as (•)w and (•)b respectively. For the sake of brevity, we assume that all onboard sensors are calibrated and are attached to the body frame. The main state of the MAV is defined as:
x=[pw,φw,{dot over (p)}b,bab,bωb,bzw]T
where pw=[xw, yw, zw]T is the 3D position in the world frame, Φw=[ψw, θw, φw]T is the yaw, pitch, and roll Euler angles that represent the 3-D orientation of the body in the world frame, from which a matrix Rwb that represent the rotation of a vector from the body frame to the world frame can be obtained. {dot over (p)}b is the 3D velocity in the body frame. bab and bωb are the bias of the accelerometer and gyroscope, both expressed in the body frame. bzw models the bias of the laser and/or pressure altimeter in the world frame.
We consider an IMU-based state propagation model:
ut=[ab,ωb]T
vt=[va,vω,vb
xt+1=ƒ(xt,ut,vt) (1)
where u is the measurement of the body frame linear accelerations and angular velocities from the IMU. vt˜N(0,Dt)εm is the process noise. va and vω represent additive noise associated with the gyroscope and the accelerometer. vba, vbω, vbz model the Gaussian random walk of the gyroscope, accelerometer and altimeter bias. The function f(•) is a discretized version of the continuous time dynamical equation [6].
Exteroceptive sensors are usually used to correct the errors in the state propagation. Following [11], we consider measurements as either being absolute or relative, depending the nature of underlying sensor. We allow arbitrary number of either absolute or relative measurement models.
A. Absolute MeasurementsAll absolute measurements can be modeled in the form:
zt+m=ha(xt+mnt+m) (2)
where nt+m˜N(0, Qt)εp is the measurement noise that can be either additive or not. ha(•) is in general a nonlinear function. An absolute measurement connects the current state with the sensor output. Examples are shown in Sect. V-B.
B. Relative MeasurementsA relative measurement connects the current and the past states with the sensor output, which can be written as:
zt+m=hr(xt+m,xt,nt+m) (3)
The formulation accurately models the nature of odometry-like algorithms (Sect. V-C and Sect. V-D) as odometry measures the incremental changes between two time instants of the state. We also note that, in order to avoid temporal drifting, most state-of-the-art laser/visual odometry algorithms are keyframe based. As such, we allow multiple future measurement (mεM, |M|>1) that corresponds to the same past state xt.
IV. UKF-Based Multi-Sensor FusionWe wish to design a modular sensor-fusion filter that is easily extensible even for inexperienced users. This means that amount of coding and mathematical deviation for the addition/removal of sensors should be minimal. One disadvantage of the popular EKF-based filtering framework is the requirement of computing the Jacobian matrices, which is proven to be difficult and time consuming for a complex MAV system. As such, we employ the derivative-free UKF based approach [1]. The key of UKF is the approximation of the propagation of Gaussian random vectors through nonlinear functions via the propagation of sigma points. Let x˜N({circumflex over (x)},Pxx)εn and consider the nonlinear function:
y=g(x), (4)
and let:
χ=[{circumflex over (x)},{circumflex over (x)}±(√{square root over ((n+λ)Pxx)})i] for i=1, . . . ,n
[i=g(χi), (5)
where g(•) is a nonlinear function, λ is a UKF parameter. (√{square root over ((n+λ)Pxx)})i is the ith column of the square root covariance matrix; which is usually computed via Cholesky decomposition. And χ are called the sigma points. The mean, covariance of the random vector y, and the cross-covariance between x and y, can be approximated as:
where ωim and ωic are weights for the sigma points. This unscented transform can be used to keep track of the covariance in both the state propagation and measurement update, thus avoiding the need of Jacobian-based covariance approximation.
A. State Augmentation for Multiple Relative MeasurementsSince a relative measurement depends both the current and past states, it is a violation of the fundamental assumption in the Kalman filter that the measurement should only depend on the current state. One way to deal with this is through state augmentation [11], where a copy of the past state is maintained in the filter. Here we present an extension of [11] to handle arbitrary number of relative measurement models with the possibility that multiple measurements correspond to the same augmented state. Our generic filtering framework allows convenience setup, addition and removal of absolute and relative measurement models.
Note that a measurement may not affect all components in the state x. For example, a visual odometry only affects the 6-DOF (Degree of Freedom) pose, not the velocity or the bias terms. We define the ith augmented state as xiεn
The addition of a new augmented state xl+1 can be done by:
Similarly, the removal of an augmented state xj is given as:
where a=n+Σi=1j-1ni and b=Σi=j+1lni. The updated augmented state covariance is given as:
{hacek over (P)}±=M±{hacek over (P)}M±T.
The change of keyframes in a odometry-like measurement model is simply the removal of an augmented state xi followed by the addition of another augmented state with the same Bi. Since we allow multiple relative measurements that correspond to the same augmented state, contrast to [11], augmented states are not deleted after measurement updates (Sect. IV-D).
This state augmentation formulation works well in an EKF setting, however, it poses issues when we try to apply it to the UKF. Since the addition of a new augmented state (8) is essentially a copy of the main state. The resulting covariance matrix {hacek over (P)}+ will not be positive definite, and the Cholesky decomposition (5) for state propagation will fail (non-unique). We now wish to have something that is similar to the Jacobian matrices for EKF, but without explicitly computing the Jacobians.
B. Jacobians for UKFIn [16], the authors present a new interpretation of the UKF as a Linear Regression Kalman Filter (LRKF). In LRKF, we seek to find the optimal linear approximation y=Ax+b+e of the nonlinear function (4) given a weighted discrete (or sigma points (6)) representation of the distribution N({circumflex over (x)},Pxx). The objective is to find the regression matrix A and vector b that minimize the linearization error e:
As shown in [16], the optimal linear regression is given by:
A=PyxPxx
The linear regression matrix A in (9) serves as the linear approximation of the nonlinear function (4). It is similar to the Jacobian in the EKF formulation. As such, the propagation and update steps in UKF can be performed in a similar fashion as EKF.
C. State PropagationObserving the fact that during state propagation only the main state changes, we start off by partitioning the augmented state and the covariance (7) into:
The linear approximation of the nonlinear state propagation (1), applied on the augmented state (7), is:
from which we can see that the propagation of the full augmented state is actually unnecessary since the only nontrivial regression matrix corresponds to the main state. We can propagate only the main state x via sigma points generated from Pt|txx and use the UKF Jacobian Ft to update the cross covariance Pt|txx
Since the process noise is not additive, we augment the main state with the process noise and generate sigma points from:
The state is then propagated forward by substituting (11) into (1), (5) and (6). We obtain {circumflex over (x)}t+1|t, the estimated value of x at time t+1 given the measurements up to t, as well as Pt+1|txx and Pt+1|tx
Pt+1|tx
The propagated augmented state and its covariance is updated according to (10):
Let there be m state propagations between two measurements, and we maintain {hacek over (x)}t+m|t and {hacek over (P)}t+m|t as the newest measurement arrives. Consider a relative measurement (3) that depends on the jth augmented state, the measurement prediction and its linear regression approximation can be written as:
Again, since only the main state and one augmented state are involved in each measurement update, we can construct another augmented state together with the possibly non-additive measurement noise:
After the state propagation (12), {grave over (P)}t+m|t is guaranteed to be positive definite, thus it is safe to perform sigma point propagation as in (5) and (6). We obtain {circumflex over (z)}t+m|t, Pt+m|tzz, Pt+m|tz{grave over (x)}, and:
Pt+m|tz{grave over (x)}{grave over (P)}t+m|t−1=[Ht+m|tx,Ht+m|txj,Lt+m].
We can apply the measurement update similar to an EKF:
{hacek over (K)}t+m={hacek over (P)}t+m|tHt+m|tTPt+m|tzz-1
{hacek over (x)}t+m|t+m={hacek over (x)}t+m|t+{hacek over (K)}t+m(zt+m−{hacek over (z)}t+m|t)
{hacek over (P)}t+m|t+m={hacek over (P)}t+m|t−{hacek over (K)}t+mHt+m|t{hacek over (P)}t+m|t
where zt+m, is the actual sensor measurement. Both the main and augmented states will be corrected during measurement update. We note that entries in Ht+m|t that correspond to inactive augmented states are zero. This can be utilized to speed up the matrix multiplication.
The fusion of absolute measurements can simply be done by {circumflex over (x)}j
As shown in
When fusing multiple measurements, it is possible that the measurements arrive out-of-order to the filter, that is, a measurement that corresponds to an earlier state arrives after the measurement that corresponds to a later state. This violates the Markov assumption of the Kalman filter. Also, due to the sensor processing delay, measurements may run behind the state propagation.
We address these two issues by storing measurements in a priority queue, where the top of the queue corresponds to the oldest measurement. A pre-defined a maximum allowable sensor delay td of 100 ms was set for our MAV platform. Newly arrived measurements that corresponded to a state older than td from the current state (generated by state propagation) are directly discarded. After each state propagation, we check the queue and process all measurements in the queue that are older than td. The priority queue essentially serves as a measurement reordering mechanism (
As the vehicle moves through the environment, global pose measurements from GPS and magnetometer may be available. It is straightforward to fuse the GPS as a global pose measurement and generate the optimal state estimate. However, this may not be the best for real-world applications. A vehicle that operates in a GPS-denied environment may suffer from accumulated drift. When the vehicle gains GPS signal, as illustrated in
This is not a new problem and it has been studied for ground vehicles [17] under the term of local frame-based navigation. However, [17] assumes that a reasonably accurate local estimate of the vehicle is always available (e.g. wheel odometry). This is not the case for MAVs since the state estimate with only the onboard IMUs drifts away vastly within a few seconds. The major difference between an IMU and the wheel odometry is that an IMU drifts temporally, but the wheel odometry only drifts spatially. However, we have relative exteroceptive sensors that are able to produce temporally drift-free estimates. As such, we only need to deal with the case that all relative exteroceptive sensors have failed. Therefore, our goal is to properly transform the global GPS measurement into the local frame to bridge the gap between relative sensor failures.
Consider a pose-only graph SLAM formulation with sk=[xkw,ykw,ψkw]Tεθ being 2D poses. We try find the optimal configuration of the pose graph given incremental motion constraints dk from laser/visual odometry, spatial loop closure constraints 1k, and absolute pose constraints zk from GPS:
The optimal pose graph configuration can be found with available solvers [18], as shown in
The pose graph SLAM provides the transformation between the non-optimized sk-1 and the SLAM-optimized sk-1+ state. This transform can be utilized to transform the global GPS measurement to be aligned with sk-1:
Δt−1=sk-1⊖sk-1+
zk-1−=Δt−1⊕zk-1
where ⊕ and ⊖ are pose compound operations as defined in [19]. The covariance Pt−1Δ of Δt−1 and subsequently the covariance Pt−1z
However, despite the large scale in our field experiments (Sect. VI), we hardly find a case that the accumulated drift is large enough to cause issues with direct GPS fusion. In the future, we will seek for even larger scale experiments to verify the necessity of the above local frame-based approach.
V. Implementation Details A. Experimental PlatformThe experimental platform shown in
Some onboard sensors are capable of producing absolute measurements (Sect. 111-A), here are their details:
1) GPS And Magnetometer:
zt=ztw+bztw+nt
If the MAVs is near hover or moving at approximately constant speed, we may say that the accelerometer output provides a pseudo measurement of the gravity vector. Let g=[0, 0, g]T, we have:
zt=RbwTgw+batb+nt.
We utilize the laser-based odometry that we developed in our earlier work [4]. Observing that man-made indoor environments mostly contains vertical walls, we can make a 2.5-D environment assumption. With this assumption, we can make use of the onboard roll and pitch estimates to project the laser scanner onto a common ground plane. As such, 2D scan matching can be utilized to estimate the incremental horizontal motion of the vehicle. We keep a local map to avoid drifting while hovering.
where P2dt=[xtw,ytw,ψtw]T,⊕2d and ⊖2d are the 2-D pose compound operations as defined in [19].
D. Relative Measurement—Visual OdometryWe implemented a classic keyframe-based visual odometry algorithm. Keyframe-based approaches have the benefit of temporally drift-free. We choose to use light-weight corner features but run the algorithm at a high-rate (25 Hz). Features are tracked across images via KLT tracker. Given a keyframe with a set of triangulated feature points, we run a robust iterative 2D-3D pose estimation [8] to estimate the 6-DOF motion of the vehicle with respect to the keyframe. New keyframes are inserted depending on the distance traveled and the current number of valid 3D points.
To achieve stable flight across different environments with possibly large orientation changes, we choose to use a position tracking controller with a nonlinear error metric [20]. The 100 Hz filter output (Sect. IV) is used directly as the feedback for the controller. In our implementation, the attitude controller runs at 1 kHz on the ARM processor on the MAV's AutoPilot board, while the position tracking control operates at 100 Hz on the main computer. We implemented both setpoint trajectory tracking and velocity control to allow flexible operations.
VI. Experimental ResultsMultiple experiments are conducted to demonstrate the robustness of our system. We begin with an quantitative evaluation in a lab environment equipped with a motion capture systems. We then test our system in two real-world autonomous flight experiments, including an industrial complex and a tree-lined campus.
A. Evaluation of Estimator PerformanceWe would like to push the limits of our onboard estimator. Therefore, we have a professional pilot to aggressively fly the quadrotor with a 3.5 m/s maximum speed and large attitude of up to 40°. The onboard state estimates are compared the ground truth from the motion capture system. Since there is no GPS measurement indoors, our system relies on a fusion of relative measurements from laser and vision. We do observe occasional laser failure due to large attitude violating the 2.5-D assumption (Sect. V-C). However, the multi-sensor filter still tracks the vehicle state throughout (
We tested our system in a challenging industrial complex. The testing site spans a variety of environments, including outdoor open space, densely filled trees, cluttered building area, and indoor environments (
We also conduct experiments in a tree-lined campus environment, as shown in
In this disclosure, we present a modular and extensible approach to integrate noisy measurements from multiple heterogeneous sensors that yield either absolute or relative observations at different and varying time intervals. Our approach generates high rate state estimates in real-time for autonomous flight. The proposed approach runs onboard our new 1.9 kg MAV platform equipped with multiple heterogeneous sensors. We demonstrate the robustness of our framework in large-scale, indoor and outdoor autonomous flight experiments that involves traversal through an industrial complex and a tree-lined campus.
In the near future, we would like to integrate higher level planning and situational awareness on our MAV platform to achieve fully autonomous operation across large-scale complex environments.
The disclosure of each of the following references is incorporated herein by reference in its entirety.
- [1] S. J. Julier and J. K. Uhlmann, “A new extension of the kalman filter to nonlinear systems,” in Proc. of SPIE, I. Kadar, Ed., vol. 3068, July 1997, pp. 182-193.
- [2] R. V. D. Merwe, E. A. Wan, and S. I. Julier, “Sigma-point kalman filters for nonlinear estimation: Applications to integrated navigation,” in Proc. of AIAA Guidance, Navigation, and Controls Conf., Providence, R.I., August 2004.
- [3] A. Bachrach, S. Prentice, R. He, and N. Roy, “RANGE-robust autonomous navigation in gps-denied environments,” J. Field Robotics, vol. 28, no. 5, pp. 644 666, 2011.
- [4] S. Shen, N. Michael, and V. Kumar, “Autonomous multi-floor indoor navigation with a computationally constrained MAV,” in Proc. of the IEEE Intl. Conf on Robot. and Autom., Shanghai, China, May 2011, pp. 20-25.
- [5] S. Weiss, M. W. Achtelik, S. Lynen, M. Chli, and R. Siegwart, “Real-time onboard visual-inertial state estimation and self-calibration of mays in unknown environments,” in Proc. of the IEEE Intl. Conf on Robot. and Autom., Saint Paul, Minn., May 2012, pp. 957-964.
- [6] D. G. Kottas, J. A. Hesch, S. L. Bowman, and S. I. Roumeliotis, “On the consistency of vision-aided inertial navigation,” in PrOC. of the Intl. Sym. on Exp. Robot., Quebec, Canada, June 2012.
- [7] F. Fraundorfer, L. Heng, D. Honegger, G. H. Lee, L. Meier, P. Tanskanen, and M. Pollefeys, “Vision-based autonomous mapping and exploration using a quadrotor MAV,” in Proc. of the IEEE/RSJ Intl. Conf on bztell. Robots and Syst., Vilamoura, Algarve, Portugal, October 2012.
- [8] K. Schmid, T. Tornio, E Ruess, H. Hirsclunuller, and M. Suppa, “Stereo vision based indoor/outdoor navigation for flying robots,” in Proc. of the IEEE/RSJ Intl. Cozzi: on Intell. Robots and Syst., Tokyo, Japan, November 2013.
- [9] A. S. Huang, A. Bachrach, P. Henry, M. Krainin, D. Maturana, D. Fox, and N. Roy, “Visual odometry and mapping for autonomous flight using an RGB-D camera,” in Proc. of the Intl. Spit. of Robot. Research, Flagstaff, Ariz., August 2011.
- [10]T. Tomic, K. Schmid, P. Lutz, A. Domel, M. Kassecker, E. Mair, I. L. Grixa, F Ruess, M. Suppa, and D. Burschka, “Autonomous UAV: Research platform for indoor and outdoor urban search and rescue,” IEEE Robot. Autom. Mag., vol. 19, no. 3, pp. 46-56, 2012.
- [11]S. I. Roumeliotis and J. W. Burdick, “Stochastic cloning: A generalized framework for processing relative state measurements,” in Proc. of the IEEE Intl. Conf on Robot. and Autom., Washington, D.C., May 2002, pp. 1788-1795.
- [12] J. Carlson, “Mapping large urban environments with GPS-aided SLAM,” Ph.D. dissertation, CMU, Pittsburgh, Pa., July 2010.
- [13]D. Schleicher, L. M. Bergasa, M. Ocaa, R. Barea, and E. Lopez, “Real-time hierarchical GPS aided visual SLAM on urban environments,” in Proc. of the IEEE Intl. Conf. on Robot. and Autom., Kobe, Japan, May 2009, pp. 4381-4386.
- [14]S. Shen, Y. Mulgaonkar, N. Michael, and V. Kumar, “Vision-based state estimation and trajectory control towards high-speed flight with a quadrotor,” in Proc. of Robot.: Sci. and Syst., Berlin, Germany, 2013.
- [15]S. Lynen, M. W. Achtelik, S. Weiss, M. Chli, and R. Siegwart, “A robust and modular multi-sensor fusion approach applied to may navigation,” in Proc. of the IEEE/RSJ Intl. Conf. on Intell. Robots and Syst., Tokyo, Japan, November 2013.
- [16]T. Lefebvre, H. Bruyninckx, and J. D. Schuller, “Comment on “a new method for the nonlinear transformation of means and covariances in filters and estimators”,” IEEE Trans. Autom. Control, vol. 47, no. 8, pp. 1406-1409, 2002.
- [17] D. C. Moore, A. S. Huang, M. Walter, and E. Olson, “Simultaneous local and global state estimation for robotic navigation,” in Proc. of the IEEE Intl. Conf. on Robot. and Autom., Kobe, Japan, May 2009, pp. 3794-3799.
- [18]R. Kuemmerle, G. Grisetti, H. Strasdat, K. Konolige, and W. Burgard, “g2o: A general framework for graph optimizations,” in Proc. of the IEEE Intl. Conf. on Robot. and Autom., Shanghai, China, May 2011, pp. 3607-3613.
- [19]R. Smith, M. Self, and P. Cheeseman, “Estimating uncertain spatial relationships in robotics,” in Proc. of the IEEE Intl. Conf on Robot. and Autom., vol. 4, Rayleigh, N.C., March 1987, p. 850.
- [20] T. Lee, M. Leoky, and N. McClamroch, “Geometric tracking control of a quadrotor uav on SE(3),” in Proc. of the Intl. Conf. on Decision and Control, Atlanta, Ga., December 2010, pp. 5420-5425.
As stated above, an autonomous rotorcraft MAV according to an embodiment of the subject matter described herein may include a trajectory generator or estimator 124 for generating a trajectory plan for controlling a trajectory of a rotorcraft MAV during flight based on an estimated current state of the rotorcraft MAV and a waypoint input by a user. The following description illustrates trajectory planning that may be performed by trajectory generator or estimator 124 according to one embodiment of the subject matter described herein.
Vision-Based Autonomous Navigation in Complex Environments with a QuadrotorThe subject matter described herein includes present a system design that enables a light-weight quadrotor equipped with only forward-facing cameras and an inexpensive IMU to autonomously navigate and efficiently map complex environments. We focus on robust integration of the high rate onboard vision-based state estimation and control, the low rate onboard visual SLAM, and online planning and trajectory generation approaches. Stable tracking of smooth trajectories is achieved under challenging conditions such as sudden waypoint changes and large scale loop closure. The performance of the proposed system is demonstrated via experiments in complex indoor and outdoor environments.
I. IntroductionQuadrotor micro-aerial vehicles (MAVs) are ideal platforms for surveillance and search and rescue in confined indoor and outdoor environments due to their small size and superior mobility. In such missions, it is essential that the quadrotor be autonomous to minimize operator workload. In this work, we are interested in pursuing a light-weight, off-the-shelf quadrotor to autonomously navigate complex unknown indoor and outdoor environments using only onboard sensors with the critical control computations running in real-time onboard the robot.
The problem of autonomous aerial navigation has been studied extensively over the past few years. Early works [1]-[3] primarily rely on laser scanners as the main sensor and localize the vehicle in indoor environments with structural elements that do not vary greatly along the vertical direction (the 2.5D assumption). Mechanized panning laser scanners that add considerable payload mass are used in [4, 5] for state estimation. Vision-based approaches, such as those in [6]-[8], rely on a downward-facing camera, a combination of stereo vision and a downward-facing optical flow sensor, and an RGB-D sensor, respectively, to achieve stable autonomous flight in indoor and/or outdoor environments. However, these approaches are unable to exploit the mobility and maneuverability of the quadrotor platform due to pragmatic concerns that arise from environment structure assumptions, reduced algorithm update rates, or the large vehicle size. Moreover, approaches that rely on downward-facing vision sensors [6, 7] often fail to perform robustly in environments with featureless floors or at low altitudes.
At the other end of the spectrum, there are many successful reactive navigation approaches that do not rely on metric state estimation [9, 10]. Although these approaches enable autonomous flight with low computation power, they fundamentally limit the flight capabilities of the MAV when operating in complex environments.
We pursue an autonomous navigation approach that enables the vehicle to estimate its state in an unknown and unstructured environment, map the environment, plan in the map, and autonomously control along trajectories developed from this plan. Online obstacle detection and replanning permit operation in static and dynamic environments with average flight speeds of more than 1 m/s. At such speeds, a low-latency state estimation, online smooth trajectory generation, and responsive vehicle control become necessary due to the agility of the platform. A challenge that arises in pursuit of this goal is the need to ensure that the estimated pose remains smooth and consistent, even during loop closures resulting from simultaneous localization and mapping. Traditionally, loop closure corrections are fused directly into the high rate onboard state estimator. This causes discontinuities in the estimated state, which, especially during rapid maneuvers, can lead to catastrophic crashes of the quadrotor.
In this work, we address these requirements by proposing a system architecture that employs two forward-facing cameras as the primary sensors, and a novel methodology that maintains estimation smoothness and control stability during replanning and loop closure, which in turn enables efficient autonomous navigation in complex environments.
II. System Design and MethodologyWe begin by providing an overview of the system architecture and methodology, and the hardware and software components required for our design. Detailed discussion of the major components are given in subsequent sections following the logical flow of the system block diagram (
A. Hardware Platform
The experimental platform (
B. Software Architecture and Methodology
The software architecture is shown in
A. Vision-Based Pose Estimation
We use a modification of our earlier work [11] to estimate the 6-DOF pose of the vehicle. Note that although we equip the platform with two cameras, we do not perform traditional stereo-based state estimation. In fact, we set one camera that captures images at 20 Hz as the primary camera, while the other camera is configured to capture images at 1 Hz. Because we don't perform high rate disparity computations, the required computational power is reduced. However, the stereo geometry allows us to estimate metric information preserving the scale of the local map and the pose estimates.
I) Monocular-Based Pose Estimation:
For images captured by the primary fisheye camera, we detect FAST corners [12] and track them using the KLT tracker [13]. Note that due to the high frame rate of the primary camera, we are able to perform feature tracking directly on the distorted fisheye images, avoiding additional computation overhead on image undistortion. We utilize the incremental rotation estimate from short term integration of the gyroscope measurement and perform 2-point RANSAC to reject tracking outliers. We propose a decoupled orientation and position estimation scheme in order to make use of distant features that are not yet triangulated. The orientation of the robot Rj is estimated via epipolar constraints with look-back history to minimize drifting. Assuming the existence of a perfect 3D local map, which consists of triangulated 3D features pi, iεI, the position of the robot rj can be found efficiently by solving the following linear system:
where uij is the unit length feature observation vector; uijrRjuij; and di=∥rj-1−pi∥.
Once the 6-DOF pose is found, the location of the feature pi can be found by solving the following linear system:
Aijpi=bij (2)
where Aij and bij represent all observations of the ith feature up to the jth frame. This is a memoryless problem, therefore the complexity of feature triangulation is constant regardless of the number of observations of that particular feature.
2) Stereo-Based Scale Recovery:
The pose estimation approach described above suffers from scale drift due to the accumulated error in the monocular-based triangulation. Every instant stereo measurement is used for scale drift compensation. Let K denote the set of features seen by both cameras. We can compute the difference of the average scene depth as:
where pis is the 3D feature location obtained solely via stereo triangulation. We can then compensate for the drifting of scale by modifying bij as (4) and solve (2) again.
B. UKF-Based Sensor Fusion
The 20 Hz pose estimate from the vision system alone is not sufficient to control the robot. An UKF with delayed measurement compensation is used to estimate the pose and velocity of the robot at 100 Hz [14]. The system state is defined as:
x=[r,{dot over (r)},q,ab]t (5)
where r is the 3D position of the robot; q is the quaternion representation the 3D orientation of the robot; and ab is the bias of the accelerometer measurement in the body frame. We use a conventional IMU-based process model to propagate the system state, and a linear measurement model which consists of the 6-DOF pose for state correction.
C. Performance of the Visual-Inertial State Estimator
We implement a visual SLAM module to eliminate the drift in the VINS system. Visual SLAM is a widely studied area. In small workspaces, approaches that use recursive filtering [15] or parallel tracking and mapping techniques [16] yield accurate results. Large scale mapping with monocular [17] or stereo [18] cameras are achieved using pose graph-based formulations. In our system, due to the limited onboard computation resources, limited wireless transmission bandwidth, and the accuracy of the onboard estimator, a high rate visual SLAM is both unnecessary and infeasible. Therefore, our visual SLAM module runs offboard with a maximum rate of 1 Hz. A pose graph-based SLAM back-end, together with a front-end that utilize SURF features [19] for wide baseline loop closure detection, yield robust performance at such low rates. We sparsely sample the estimated robot trajectory to generate nodes for the pose graph. For each node, we compute sparse 3D points by detecting and matching SURF features between the stereo images. Dense disparity images and dense point clouds are also computed.
We detect loop closures by checking nodes that fall inside the uncertainty ellipsoid of the current node. We check a constant number of nodes, starting from the earliest candidate, for possible loop closures. SURF features are used to test the similarity between two scenes. We compute the relative transform between the current node and the loop closure candidate using RANSAC PnP [20]. A rigidity test, proposed in (Sect. 3.4, [21]), is performed to verify the geometric consistency of the loop closure transform. Candidate transforms that pass the geometric verification are added to the pose graph. Finally, we use the iSAM library for pose graph optimization [22]. Once an optimized pose graph is found, we can construct a 3D voxel grid map by projecting the dense point cloud to the global frame. This map is used for the high level planning (Sect. V) and to enable the human operator to monitor the progress of the experiment. The optimized pose represents an estimate in the world frame and is denoted by (rjW, RjW).
The pose correction from the visual SLAM djWO, which serves as the transform between the odometry frame and the world frame, is formulated such that:
(rjW,R|jW)=djWO⊕(rjO,RjO) (6)
where ⊕ is the pose update function defined in [23]. In contrast to traditional approaches, we do not use (rjW,RjW) as a global pose measurement for correcting the drift in the VINS system. Instead, we feed djWO, into the trajectory generator (Sect. VI) and compute trajectories that are guaranteed to be smooth even if there are large discontinuities in the visual-SLAM pose estimate (i.e. ∥djWO⊖dj-1WO∥ is large) due to loop closures. This is the major departure of our system from existing approaches and it is the key to enable high-speed autonomous navigation in complex environments. Further details are provided in Sect. VI.
V. High Level PlanningWe employ a two-stage planning approach. On a higher level, given the user-specified waypoints in the world frame, and treating the quadrotor as a cylinder, a high level path that connects the current robot position and the desired goal, which consists a sequence of desired 3D positions and yaw angles, is generated using the RRT* [24] as implemented in the Open Motion Planning Library (OMPL) [25]. The resulting path is simplified to a minimum number of K waypoints gkW and is sent to the trajectory generator (Sect VI) for further refinement. The path is checked for possible collisions at the same frequency as the map update (1 Hz, Sect IV). Although the high level planner only requires moderate computational resources, we run it offboard as all information required for high level planning comes from the offboard visual SLAM module. We also allow the user to bypass the planner and explicitly set a sequence of waypoints.
VI. Trajectory GenerationWe first transform all waypoints from the high level planner into the odometry frame using the latest pose correction from the visual SLAM (6):
gkO=⊖djWO⊕gkW (7)
If the robot flies through all transformed waypoints using the state estimate in the odometry frame for feedback control, it will also fly through the same sets of waypoints in the world frame. Moreover, it there are large scale loop closures (i.e. large changes in djWO), the set of waypoints that the robot is heading towards will change significantly. However, if we are able to regenerate smooth trajectories with initial conditions equal to the current state of the robot, the transition between trajectories will be smooth and no special handling is needed within the onboard state estimator and the controller.
We wish to ensure that the quadrotor smoothly passes through all waypoints, while at the same time maintaining a reliable state estimate. A crucial condition that determines the quality of the vision-based estimate is the tracking performance. With our fisheye cameras setup, it can be seen from
By differentiating the equation of motion of a quadrotor [26], it can be seen that the angular velocity of the body frame is affinely related to the jerk, the derivative of the linear acceleration. As such, we generate trajectories that minimize the jerk of the quadrotor in horizontal directions.
For the vertical direction, we wish to minimize the RPM changes of the motors, which again correspond to the jerk. Intermediate waypoints are added shortly before and after a waypoint if the angle between the two line segments that connect this waypoint exceeds a threshold in order to avoid large deviations from the high level path. We utilize a polynomial trajectory generation algorithm [27] that runs onboard the robot with a runtime on the order of 10 ms. Optimal trajectories can be found by solving the following unconstrained quadratic programming:
Where y is a collection of desired derivative values at each waypoint, which can be either free or fixed. We fix the position, velocity, acceleration, at the first waypoint to be current state of the robot in order to maintain smooth trajectories during replanning and loop closures. The velocity and acceleration are set to be zero for the last waypoint. For all other waypoints, only position is fixed and the trajectory generator will provides the velocity and acceleration profile. The coefficients of the polynomial trajectories s can be found via a linear mapping s=My.
A limitation of the above trajectory generation approach is the necessity of predefining the travel time between waypoints. Due to computational constraints, we do not perform any iterative time optimization [27, 28] to find the optimal segment time, but rather use a heuristic that approximates the segment time as a linear trajectory that always accelerates from and decelerates to zero speed with a constant acceleration at the beginning and end of a segment, and maintains constant velocity in the middle of a segment. This simple heuristic can help avoid excessive accelerations during short segments, and is a reasonable time approximation for long segments.
A. Position Tracking Controller
For this work, we choose to use a position tracking controller with a nonlinear error metric [29] due to its superior performance in highly dynamical motions that involve large angle changes and significant accelerations. The 100 Hz state estimate from the VINS system (Sect. III) is used directly as the feedback for the controller. In our implementation, the attitude controller runs at 1 kHz on the ARM processor on the robot's AutoPilot board, while the position tracking control operates at 100 Hz on the Atom processor.
B. Hybrid-System Controller
Although our goal is to develop a fully autonomous vehicle, at some point during the experiment, the human operator may wish to have simple, but direct control of the vehicle. As such, we developed a finite state machine-based hybrid-system controller (
We present three representative experiments to demonstrate the performance of the proposed system. The first experiment demonstrates the ability of the proposed system to maintain globally consistent tracking. We provide a comparison with ground truth to quantify the performance. In the second experiment, the robot navigates an indoor environment with a large loop (approximately 190 m) and completes the loop within one battery charge (less than 5 minutes of flight time). Finally, we present an outdoor navigation experiment that emphasizes the robustness of the proposed system against environment changes and strong wind disturbance.
A. Evaluation of System Performance with Ground Truth Comparison
In this experiment, the robot autonomously follows a smooth trajectory generated from a rectangle pattern at approximately 1 m/s. The ground truth from Vicon is used to quantify the global tracking performance. As seen from
B. Navigation of Indoor Environments with Large Loops
We now consider a case where the robot autonomously navigates through a large-scale environment with loops. Due to the size of the loop (approximately 190 m), and the short battery life cycle (less than 5 min), we must achieve high-speed navigation in order to complete the task. This environment poses significant challenges to approaches that uses downward facing cameras [6, 7] due to the featureless floor (
This experiment demonstrates the performance of the proposed system in outdoor environments. The experiment is conducted in a typical winter day at Philadelphia, Pa., where the wind speed goes up to 20 km/hr. The total travel distance is approximately 170 m with a total duration of 166 s (
As described herein, we propose a system design that enables globally consistent autonomous navigation in complex environments with a light weight, off-the-shelf quadrotor using only onboard cameras and an IMU as sensors. We address the issue of maintaining smooth trajectory tracking during challenging conditions such as sudden waypoint changes and loop closure. Online experimental results in both indoor and outdoor environments are presented to demonstrate the performance of the proposed system.
An integrated laser- and/or GPS-based state estimation approach may be incorporated into our current system to extend the operational environments and enhance the system robustness.
The disclosures of each of the following references in incorporated herein by reference in its entirety.
REFERENCES
- [1]. A. Bachrach, S. Prentice, R. He, and N, Roy, “RANGE-robust autonomous navigation in gps-denied environments,” J. Field Robotics, vol. 28, no. 5, pp, 644-666, 2011.
- [2]. S. Grzonka, G. Grisetti, and W. Burgard, “A fully autonomous indoor quadrotor,” IEEE Trans. Robot., vol. PP, no. 99, pp. 1-11, 2011.
- [3]. S. Shen, N. Michael, and V. Kumar, “Autonomous multi-floor indoor navigation with a computationally constrained MAV,” in Proc. of the IEEE Intl. Conf on Robot. and Autom., Shanghai, China, May 2011, pp. 20-25.
- [4]. S. Scherer, J. Rehder, S. Achar, H. Cover, A, Chambers, S. Nuske, and S. Singh, “River mapping from a flying robot: state estimation, river detection, and obstacle mapping,” Auton. Robots, vol. 33, no. 1-2, pp. 189-214, August 2012.
- [5]. A. Kushleyev, B. MacAllister, and M. Likhachev, “Planning for landing site selection in the aerial supply delivery,” in Proc. of the IEEE/RSJ Intl. Conf on Intell. Robots and Syst, San Francisco, Calif., September 2011, pp, 1146-1153.
- [6]. S. Weiss, M. W. Achtelik, S. Lynen, M. Chli, and R. Siegwart, “Real-time onboard visual-inertial state estimation and self-calibration of mays in unknown environments,” in Proc. of the IEEE Intl. Conf. on Robot. and Autom., Saint Paul, Minn., May 2012, pp. 957-964.
- [7]. F. Fraundorfer, L. Heng, D. Honegger, G. H. Lee, L. Meier, P. Tanskanen, and M. Pollefeys, “Vision-based autonomous mapping and exploration using a quadrotor MAV,” in Proc. of the IEEE/RSJ Intl. Conf on Intell. Robots and Syst., Vilamoura, Algarve, Portugal, October 2012.
- [8]. A. S. Huang, A. Bachrach, P. Henry, M. Krainin, D. Maturana, D. Fox, and N. Roy, “Visual odometry and mapping for autonomous flight using an RGB-D camera,” in Proc. of the Intl. Sym, of Robot. Research, Flagstaff, Ariz., August 2011,
- [9]. C. Bills, J. Chen, and A. Saxena, “Autonomous MAV flight in indoor environments using single image perspective cues,” in Proc. of the IEEE Intl. Conf. on Robot, and Autom, Shanghai, China, May 2011, pp. 5776-5783.
- [10]. G. de Croon, C. D. Wagterb, B. Remesb, and R. Ruijsinkb, “Sub-sampling: Real-time vision for micro air vehicles,” Robot. and Autom. Syst., vol. 60, no. 2, pp. 167-181, February, 2012.
- [11]. S. Shen, Y. Mulgaonlcar, N. Michael, and V. Kumar, “Vision-based state estimation for autonomous rotorcraft MAVs in complex environments,” in Proc. of the IEEE Intl. Conf on Robot. and Autom, Karlsruhe, Germany, May 2013, To appear.
- [12]. E. Rosten and T. Drummond, “Machine learning for high-speed corner detection,” in Proc. of the European Conf on Computer Vision, Graz, Austria, May 2006.
- [13]. B. D. Lucas and T. Kanade, “An iterative image registration technique with an application to stereo vision,” in Proc. of the Intl. Joint Conf on Artificial Intelligence, Vancouver, Canada, August 1981, pp. 24-28.
- [14]. R. V. D. Merwe, E. A. Wan, and S. I. Julier, “Sigma-point Kalman filters for nonlinear estimation: Applications to integrated navigation,” in Proc. of AIAA Guidance, Navigation, and Controls Conf, Providence, R.I., August 2004.
- [15]. J. Civera, A. J. Davison, and J. Montiel, “Inverse depth parameteriza-tion for monocular SLAM,” IEEE Trans. Robot, vol. 24, no, 5, pp. 932-945, October 2008.
- [16]. G. Klein and D. Murray, “Parallel tracking and mapping for small AR workspaces,” in Proc. Sixth IEEE and ACM International Symposium on Mixed and Augmented Reality (ISMAR'07), Nara, Japan, November 2007.
- [17]. H. Strasdat, J. M. M. Montiel, and A. J. Davison, “Scale drift-aware large scale monocular SLAM,” in Proc. of Robot.: Sci. and Syst., Zaragoza, Spain, June 2010.
- [18]. C. Mei, G. Sibley, M. Cummins, P. Newman, and I. Reid, “RSLAM: A system for large-scale mapping in constant-time using stereo,” Intl J. of Computer Vision, pp. 1-17, June 2010.
- [19]. H. Bay, T. Tuytelaars, and L. V. Gool, “SURF: Speeded up robust features,” in Proc. of the European Conf on Computer Vision, Graz, Austria, May 2006.
- [20]. F. Moreno-Noguer, V. Lepetit, and P. Fua, “Accurate non-iterative 0(n) solution to the PnP problem,” in Proc. of the IEEE Intl. Conf on Computer Vision, Rio de Janeiro, Brazil, October 2007.
- [21]. E. B. Olson, “Robust and efficient robotic mapping,” Ph.D. dissertation, MIT, Cambridge, Mass., June 2008.
- [22]. M. Kaess, A. Ranganathan, and F. Dellaert, “iSAM: Incremental smoothing and mapping,” IEEE Trans, Robot., vol. 24, no. 6, pp. 1365-1378, December 2008.
- [23]. R. Smith, M. Self, and P. Cheeseman, “Estimating uncertain spatial relationships in robotics,” in Proc. of the IEEE Intl. Conf on Robot. and Autom., Rayleigh, N.C., March, 1987, p. 850.
- [24]. S. Karaman and E. Frazzoli, “Incremental sampling-based algorithms for optimal motion planning,” in Proc. of Robot: Sci, and Syst., Zaragoza, Spain, June 2010.
- [25]. I. A. Sucan, M. Moll, and L. E. Kavraki, “The Open Motion Planning Library,” IEEE Robot. Autom. Mag., vol. 19, no. 4, pp. 72-82, December 2012, http://ompl.kavrakilab.org.
- [26]. N. Michael, D. Mellinger, Q. Lindsey, and V. Kumar, “The GRASP multiple micro UAV testbed,” IEEE Robot, Autoin. Mag., vol. 17, no. 3, pp. 56-65, September, 2010,
- [27]. C. Richter, A. Bry, and N. Roy, “Polynomial trajectory planning for quadrotor flight,” in Proc. of the IEEE Intl, Conf on Robot, and Autom, Karlsruhe, Germany, May 2013, To appear.
- [28]. D. Mellinger and V. Kumar, “Minimum snap trajectory generation and control for quadrotors,” in Proc. of the IEEE Intl, Conf on Robot, and Autom, Shanghai, China, May 2011, pp. 2520-2525.
- [29]. T. Lee, M. Leoky, and N. McClamroch, “Geometric tracking control of a quadrotor uav on SE(3),” in Proc. of the Intl. Conf. on Decision and Control, Atlanta, Ga., December 2010, pp. 5420-5425.
The subject matter described herein includes any combination of the elements or techniques described herein even if not expressly described as a combination. For example, elements or methodologies described in the section entitled Vision Based Autonomous Navigation in Complex Environments with a Quadrotor can be combined with any of the methods or elements described prior to that section.
It will be understood that various details of the subject matter described herein may be changed without departing from the scope of the subject matter described herein. Furthermore, the foregoing description is for the purpose of illustration only, and not for the purpose of limitation.
Claims
1. A system that enables autonomous control of a vehicle in indoor and outdoor environments, the system comprising:
- a sensor fusion module for combining measurements from a plurality of sensors of different modalities to estimate a current state of the vehicle given current and previous measurements from the sensors and a previous estimated state of the vehicle, wherein the sensor fusion module is configured to maintain smoothness in the state estimates of the vehicle when: one or more sensors provide inaccurate information, when global positioning system (GPS) measurements are unavailable after a period of availability, or when GPS measurements become available after a period of unavailability; and
- a trajectory generator for generating a plan for controlling a trajectory of the vehicle based on the estimated current state and a goal or a waypoint input provided by either a user or a higher level planner.
2. The system of claim 1 wherein the sensors include an inertial measurement unit (IMU), and at least one of a pressure altimeter, a magnetometer, a laser scanner, a camera, a downward facing optical sensor, and a global positioning system (GPS) receiver.
3. The system of claim 1 wherein the sensor fusion module is configured to use an Unscented Kalman Filter (UKF) to combine the measurements from the sensors of different modalities, enabling addition and removal of sensors without reconfiguration of software of the sensor fusion module.
4. The system of claim 3 wherein the sensor fusion module is configured to estimate the current state using current relative measurements and copies of augmented past states in the filter.
5. The system of claim 3 wherein the sensor fusion module is configured to judiciously remove augmented states from the filter and add new augmented states to the filter.
6. The system of claim 3 wherein the sensor fusion module is configured to fuse measurements from the sensors that arrive out of order to the filter.
7. A method that enables autonomous control of a vehicle in indoor and outdoor environments, the method comprising:
- combining measurements from a plurality of sensors of different modalities to generate an estimate of a current state of the vehicle given current measurements from the sensors and a previous estimated state of the vehicle;
- generating a signal for planning a trajectory of the vehicle based on the estimated current state and a goal or waypoint input by a user or a higher level planner; and
- smoothing changes in state of the vehicle when: output from one or more of the sensors is inaccurate, global positioning system (GPS) measurements become available after a period of unavailability, or GPS measurements become unavailable after a period of availability.
8. The method of claim 7 wherein the sensors include at least an inertial measurement unit (IMU) and at least one of a pressure altimeter, a magnetometer, a laser scanner, a camera, and a GPS receiver.
9. The method of claim 7 wherein combining the measurements includes an Unscented Kalman Filter (UKF) to combine the measurements from the sensors of different modalities, enabling addition and removal of sensors without reconfiguration of the sensor fusion module.
10. The method of claim 9 wherein estimating the current state includes using current relative measurement and copies of augmented past states in the filter.
11. The method of claim 9 comprising removing augmented states from the filter in response to addition of a new augmented state with a binary selection matrix corresponding to that of a previous augmented state.
12. The method of claim 9 comprising fusing measurements from the sensors that arrive out of order at the filter.
13. A non-transitory computer readable medium having stored thereon executable instructions that when executed by the processor of a computer controls the computer to perform steps comprising:
- combining measurements from a plurality of sensors of different modalities to generate an estimate of a current state of the vehicle given current measurements from the sensors and a previous estimated state of the vehicle;
- generating a signal for planning a trajectory of the vehicle based on the estimated current state and a goal or waypoint input by a user or a higher level planner; and
- smoothing changes in state of the vehicle when: output from one or more of the sensors is inaccurate, global positioning system (GPS) measurements become available after a period of unavailability, or GPS measurements become unavailable after a period of availability.
14. The system of claim 1 wherein the vehicle comprises a rotorcraft micro-aerial vehicle (MAV).
15. The method of claim 7 wherein the vehicle comprises a rotorcraft micro-aerial vehicle (MAV).
16. The non-transitory computer readable medium of claim 13 wherein the vehicle comprises a rotorcraft micro-aerial vehicle (MAV).
Type: Application
Filed: May 26, 2016
Publication Date: Jul 27, 2017
Inventors: R. Vijay Kumar (Wilmington, DE), Shaojie Shen (Philadelphia, PA), Nathan Michael (Munhall, PA), Kartik Mohta (Philadelphia, PA)
Application Number: 15/165,846