Efficient Autonomous navigation for planetary rovers with limited resources (Paper Review)-Part 2

Paper Name:- Efficient autonomous navigation for planetary rovers with limited resources

Author Name:- Levin Gerdes, Martin Azkarat, José Ricardo Sánchez‐Ibáñez, Luc Joudrier, Carlos Jesús Perez‐del‐Pulgar

DOI:- 10.1002/rob.21981

(Please refer to the paper for the reference that I have mentioned in the review and want to know more about the topic)

Note:- This is a second part of the paper review if you haven't read the first part of it. I suggest you should first part first.

Building Blocks for the efficient autonomous navigation

The authors describe the three key components they created for achieving the proposed efficient autonomous navigation mode with hazard avoidance capabilities. The trajectory control, hazard identification, and local path planner are the components in question. Our objective is to present a realistic approach for speedier autonomous navigation for Mars planetary rovers on terrains of low to moderate difficulty. While the approach presented here is interesting for the ExoMars mission, these characteristics may become even more important in future Mars exploration missions, such as the Sample Fetch Rover of the Mars Sample Return program, where the travel distance and speed requirements are significantly higher.

Using the Rover Exomars Case using the Rover Exomars Case, the proposed control architecture in this article uses the localization chamber (LOCCAM) and the error feature of the VO component that works on the board. We do not solve the localization subject in this study, but instead, use Loccam using Loccam using Loccam as inertia data of the Odomotry wheel encoder to accept similar implementation with one of the executives through the predetermined decisions of the estimates. Inertia measurement (IMU), is called Bio. As can be seen in FIG. 2, the path planned by the ground control operator is supplied to the local planner of the path stored in the track control component and the memory. The value of the stereo system emitted from the VO process is supplied to the danger detection module. This module constantly confirms the risk. When an obstacle is detected, a registered hazard is sent to the local route planner. Data is sent to the patency grid, which contains 1s and 0s to indicate that each area of ​​the grid is at risk. This search then runs a local route replanning job that computes a new route using the last estimated location, the original planned route, and the registered hazard information. The newly calculated path will return to the original path after bypassing the detected obstacle.

Compared to the ExoMars route that follows the exploration approach, the proposed architecture works with little computational cost, but only adds risk detection checks. The efficient risk detection algorithm described in Section 2.2 does not require much additional computation because it uses the data processed by the VO component. The local route planner is warned to start replanning only after a risk is detected. Between discoveries, the route planner waits quietly. After the triggered rescheduling is complete, the route planner resumes idle mode. An important feature of this navigation approach is that it provides fast passage and potentially continuous operation. Simple driving in driving is only necessary to avoid dangerous risks for starting. In a standard approach to independent executable navigation as for Exomars, the Rover periodically compares the front terrain by periodically using the search camera (Navcam) periodically to compare the front terrain and resumes the intersection of the digital height card (DEM) Analyzes the intersection of the digital height card (DEM) before. movement. Refusal emphasizes the effects of the difference between the calculation load and the proposed approach compared to a variety of autonomous navigation approaches. The following sections provide detailed information about the implementation of three main components.

Trajectory Control

This section describes the algorithms used to control the locus performed by the Rover. This algorithm is used to perform this path regardless of whether it is a global path of repeated paths locally to avoid the following navigation or danger. The search approach approaches from a more basic method after a more complex full self-navigation. This path consists of a list of points in the operating disassembly space called the Waypoint that the robot should inquire about. The distance between the sequential track point (segment length) of the path throughout the path is generally considered the same structure as the same structure with a single difference in space, and the global and local paths are easily considered. The work of the trajectory control module consists of the steering vehicle along the path of the last waypoint or target position. Routes are provided from the ground by higher-level components, such as onboard route planners or workers, and are assumed to be safe to travel up to a certain deviation. This introduces the concept of a nominal track consisting of a line segment connecting all waypoints along the route and a safety corridor, an area surrounded by two parallel lines to the left and right of the nominal track at a certain symmetrical distance. Tracking error is defined as the lateral deviation of the nominal path measured from the rover position to the nearest segment line.

The event of the track described in this article is based on the popular pursuit of the pursuit algorithm. This algorithm has been successfully used to control the car track of teams in the magnificent and urban work of Darpa and control the trajectory of planetary trigops. In the development of named (or conservative) algorithms, we originally approach to ensure that the path is limited in a predefined security corridor corresponding to the restriction of a predefined security corridor (ε) This maximum error is considered an important security function for the planetary exploration mission. C-PEP also generally does not generate a soft interpolation track that is typically performed in the path algorithm. Instead, it guides the rover along a path, allowing controlled cornering to transition smoothly between successive segments. Finally, since it is a geometry-based algorithm, its parameters are intuitively configurable based on the rover's physical characteristics and capabilities. The opposite of the proposed approach is a control theory-based algorithm designed as a P/PI controller and using transformation and heading errors as feedback inputs to the control law loop.

c- Pursuit

This algorithm is designed for a special focus on Rovers for planetary exploration and given the oriented conditions. Rovers generally move at low speed in the order of size cm / s. As a result, it can forget the dynamic equation, and the problem can be expressed using only the kinetic equation. We will see the Rover Exercise System if there are four steering drives for experimental rows used for experimental verification for the previous six-character wheels. The four steering drives are to provide sufficient agility that can perform motion such as ACKERMANN and follow the trajectory in the point turn, 2D (2D) plane required for execution. Ackermann maneuvers are generally more efficient than exact turns. The latter can be done while the rover needs to stop and align the rudder before turning, while the former can move forward without stopping while continuing to adjust the steering angle.

The Ackermann steering of a six‐wheeled rover with four steering wheels can be thought of as a bicycle 2D kinematics model where the bike's rear-wheel represents the center of the rover body and the bike's front wheel is in the center of the two front wheels of the rover. Given that the rover is moving on a surface, its state is represented with the 2D position coordinates p =[ ] x y, r, and its orientation (or heading) θ.

The original pure‐pursuit algorithm solves the path following problem of an Ackermann‐steered vehicle by providing the 2D motion command (translational and rotational velocities) that will drive the system along the path. Since the linear speed (body velocity) is typically given as a configuration parameter, only the turning radius (or radius of curvature) needs to be found. This fixes the rotational speed and the Ackermann maneuver to be executed is uniquely defined. The main steps of the algorithm are as follows:

1. Determine the position of the vehicle pr with respect to the nominal path.

2. Determine a look‐ahead point p(lh) ahead on the path.

3. Calculate the arc s(turn) required to steer the vehicle from p(r) to p(lh).

where the look‐ahead point is a point in the nominal path that is found at a look‐ahead distance d(lh) from the rover position in the direction of the motion. The vehicle is instantaneously steered toward this virtual point. As the vehicle moves, this process is iteratively repeated and the updated look‐ahead point also shifts forward, hence the name pure‐pursuit.

In c‐pursuit, instead of having a constant parameter for the look-ahead distance, this is dynamically adapted depending on the tracking error ϵ of the rover at each control iteration.

where d(lh) is the fixed‐parameter setting and kϵ > 0 is the gain of the deviation penalization. Lowering the value of the look‐ahead distance has the effect of bringing the rover back to the nominal path faster, with a sharper turning radius. In the extreme cases where the curvature exceeds the steering limits, a point turn is commanded to realign the rover with the heading toward the look‐ahead point. In pure‐pursuit, the look‐ahead point is found at the crossing point of the nominal path with a circle of radius equal to the look‐ahead distance centered at the rover location. In c‐pursuit, the look‐ahead point is instead at a distance d′lh along the path from the projected position p(pr), that is, the position of the nominal path closest to the rover location. Figure 3 illustrates the main steps of the c‐pursuit algorithm. The segment between two consecutive waypoints (p w,n ) is shown. The parallel green lines define the safety corridor. The circles around the waypoints delimit the area under which a waypoint is considered achieved. The rover is set at a given arbitrary location along the segment, denoted by the x y − colored frame. The figure demonstrates how the look‐ahead point is found to determine the Ackermann maneuver that steers the rover toward it. As the vehicle moves, the look‐ahead point shifts forward along the path and this control scheme results in the rover executing the trajectory shown in the dotted gray line, which gradually converges back towards the path while progressing along with it. 

Interestingly, the modifications added by c‐pursuit have a relevant impact when two connecting segments form an increasingly sharp angle. Both algorithms allow a certain deviation from the nominal path, that is, “cutting” corners, and guide the rover through smooth trajectories with Ackermann maneuvers instead of sequences of straight lines and point turns. However, pure‐pursuit can potentially guide the rover toward the outside of the safety corridor. This tendency is especially pronounced when the change in heading between segments is greater than 45*. We define here γ as the angle formed between two segments to represent this change in heading. In Figure 4, we illustrate the behavior of both algorithms. The plotted vector fields show the direction to the look‐ahead point p(lh) for different scenarios. These vector fields represent the nominal trajectories that the rover would execute in each scenario similar to the gray dotted trajectory line shown in Figure 3.

As can be seen in Figure 4b, c‐pursuit is always directing the rover “inwards” within the safety corridor, even for γ =  90*. In the same situation, the original pure‐pursuit cuts corner more aggressively, thereby violating the safety corridor constraints.

Hazard Detection

The hazard detection strategy should be computationally inexpensive and rely only on the (engineering) sensors planned to be included on the ExoMars rover.

Examples of terrain traversability and hazard detection using sensors such as time‐of‐flight (ToF) cameras or LiDARs are common in terrestrial applications these days since these are present in many autonomous navigation system solutions for cars. However, these are not suitable for space rovers due to their low TRL and high power consumption. Additionally, these sensors are not currently included on any Mars planetary rover. A recent example of a terrestrial application using optical cameras is found in Andrakhanov and Stuchkov (2017). However, it relies on machine learning techniques for terrain classification and obstacle detection, and machine learning technology has not been adopted yet by space missions due to its non-deterministic property. The majority of techniques used by main actors in space missions rely on the generation of a DEM that is later analyzed for traversability. This is a computationally demanding task since it requires the calculation and processing of three‐dimensional (3D) points in the scene. These solutions are therefore difficult to implement in a fast‐traversing continuous‐driving mode and it is usually necessary to stop the rover every few meters to calculate the next segment of the local path. The latest work in Marc and Weclewski (2019) and Marc, Weclewski, and Lachat (2018) propose the use of parallelization of tasks together with a less demanding terrain analysis over the DEM as a way to introduce a faster navigation mode in terrains of moderate difficulty. This approach is conceptually similar to the navigation mode presented in this paper but still requires the rover to stop periodically for mapping, that is, DEM generation. Back in 1998, a hazard detection technique without the need for DEM generation was presented for terrestrial applications. Instead of using DEMs, the authors fit potential ground planes to the disparities provided by a stereo camera. Hazards are detected by measuring deviations from the average height in the fitted ground plane. While this is an efficient and fast computing solution, it becomes less suitable when applied to space applications due to the difficulty of fitting a ground plane to uneven and unstructured planetary surfaces. In a similarly interesting work, investigated how different approaches to stereo calibration could be utilized for mobile robotics applications on Earth. Their work presents a computationally cheap solution for detecting hazards without the generation of DEMs. The authors propose to perform a height calibration of the camera stereo bench instead of a depth calibration. In doing so, the disparity map immediately provides the height in each pixel. However, this is again a potentially invalid solution for space applications, since if this were to be applied we would need to repurpose the LocCam's or NavCam's disparity calculations, given that these are computed according to depth calibration parameters that are needed for localization and navigation functions.

In our approach, we chose to develop the hazard detection based on the already existing stereo vision capabilities of the LocCam located on the front end of the ExoMars rover. The necessary algorithms and calibrations for computing the disparities and distances are already present since they are needed for VO. The goal is to develop hazard detection as a safety feature that is fast to compute and does not compute more than what is necessary. With this in mind, we do not compute a 3D map representation of the terrain in front of the rover nor do we need to know what the next ten meters look like. Instead, we aim to create a component that detects hazards that are right in front of the rover and allows the rover to quickly react and stop before colliding with them by implementing the “electronic bumper” which detects possible obstacles, rocks, or craters, and alerts the rover to stop the motion. This detector only searches for hazards in the close vicinity of the rover's front wheels (e.g.<0.5 m). The reason for this short‐range detector is twofold: it reduces the computational load for implementing the detector and is more flexible to allow more maneuvers, that is, the rover will not stop if the hazard is 3 m in front, and the detection is only triggered when the hazard falls inevitably onto the rover path. This feature, as simple as it may sound, is key to allowing the operators to do their tactical planning less conservatively, without increasing the risks for the rover. These requirements allow us to have a small region of interest (RoI) within the provided stereo images. In Figure 5, visualization of this RoI overlaid on top of the left camera's image is shown.

Local Path Planner(Hazard avoidance)

When a hazard is detected, the rover acquires more data describing the environment in which it is moving. The path being followed up until the detection of the hazard was previously created without considering this unknown information. This detection warns the system about the fact that the next waypoints may be located in unsafe places. Therefore, a local path planner is employed to dynamically update the path according to the newly received traversability data. In particular, we make use of the Dynamic Multilayered Path Planner (DyMu), an algorithm we designed and developed. Its main purpose is twofold: to compute an optimal global path subject to energy and/or time criteria, and to provide the ability to dynamically update it during the traverse. DyMu offers several strategies to perform the path update, and we opt for the one being most conservative. In this approach, instead of modifying the entire remaining path, it only recomputes those waypoints from the original path that are close to both the hazardous element and the rover. Figure 8 depicts an illustrative example of this functionality, where a local section of path is produced to replace the aforementioned waypoints. DyMu calculates the shape of this local section according to certain cost values given to the three areas portrayed in Figure 8.