• No results found

Recovery attempt through inverse trajectory search

As an alternative, an inverse trajectory recovery approach that takes the path traveled by the robot into account is used instead as shown in Figure 7.4. Based on the current pose where the robot is stuck Ps, the path followed by the robot so far is traced back from end to start. A threshold radius rr around Ps is used to determine the pose on a circle around the robot. This pose Pr is selected as the recovery target pose and sent to the robot controller. By following this approach, information about traversable terrain is implicitly taken into account and the robot attempts to get unstuck via a path that was already traversed before.

60 7 Navigation, Exploration, and Search for Victims

7.6 Improved Search for Victims

The goal in USAR is finding and extricating victims. Contributions detailing approaches for per-forming a more thorough search and for reducing the false position detection rate are presented in the next sections.

7.6.1 Victim Exploration

To detect objects of interest in disaster scenarios reliably, a thorough search has to be performed by (autonomous) robotic systems tasked with the exploration of the environment. As noted previously, exploration of unknown environments has been well studied in the past, but only few approaches consider different sensing modalities. Often, an environment that has been covered by a LIDAR sensor is considered explored. This approach is sufficient for tasks like learning a map of an unknown environment, but it does not take into account the limited effective detection range of sensors when searching for objects of interest, like victims in a disaster environment. The reliable detection of victims trapped under rubble using thermal imaging requires the USAR robotic system to come as close as possible to permit detection even when only very small portions of the victim are visible to sensors.

Approaches that attempt to estimate the information gain for exploration goals have been proposed [144]. With victims potentially covered by small holes in the environment, estimating potential information gain for a given pose in the environment is much more challenging than in many other scenarios, however.

While approaches for an exhaustive search for voids in volumetric environment models have been demonstrated [39], computational complexity precludes their real-time application. As an alternative approach, instead of considering parts of the environment covered by sensors explic-itly, the path planner is modified to generate paths that implicitly lead to improved coverage of the environment by relevant sensors.

This is achieved by not approaching frontier cells directly as in the original exploration trans-form approach but instead finding the closest free cell to the robot that is at least a threshold dt away from the path traveled by the robot so far.

Effectively, this leads to the robot taking much shorter steps during exploration of the envi-ronment, resulting in better coverage. While a somewhat similar behavior could be achieved by artificially reducing LIDAR sensor range, an important advantage of the proposed approach is the fact that the distance per step can be varied dynamically at runtime. It is also possibly to switch between standard and victim exploration dynamically, which would not be possible when generating reduced range maps.

Figure 7.5 shows two examples from victim search missions at RoboCup 2014. For the upper map, the standard exploration approach was used. This resulted in the robot missing most victims in the map. In contrast to this, the victim exploration approach was used for the lower map. As is visible, it makes the robot visit also smaller cavities in the environment, resulting in more victims being detected.

In Section 7.9.1, a comprehensive comparison between both approaches is described.

7.6 Improved Search for Victims 61

(a)

(b)

Figure 7.5: Examples of victim search at RoboCup 2014: (a): Use of exploration transform. The UGV explores frontiers between known and unknown space in the grid map. (b):

Exploration by using victim exploration, generating target poses close to robot path.

7.6.2 Robust Navigation towards Goal Poses

Both in autonomous and semi-autonomous operation, target poses for the robotic system might not be valid due to obstacles preventing the robot from reaching the target pose. Without careful consideration of this, the planner will report that planning failed as it is unable to plan to a given target pose. In many practical cases, however, a pose close by might be reachable, allowing continuation of the robot mission without interruption.

A common use case observed in the USAR scenario was USAR robots approaching victims for closer inspection. To reach those victims that are typically situated inside or close by to obstacles, just forwarding the victim pose to the planner as a goal pose would naturally fail.

Instead, an observation pose has to be found that is sufficiently close to the victim as to provide sensors with a view that allows inspection and possible confirmation of the victim hypothesis.

62 7 Navigation, Exploration, and Search for Victims

To achieve this, an approach generating valid observation poses is added to the planner. Based on an obstacle transform map, an area around the target pose (2m x 2m per default) is sampled for valid poses the robot can reach. Only poses that are within free space (as opposed to un-known space) are considered valid here, as allowing unun-known space can lead to wrong target locations in some degenerate cases.

The approach is shown in Algorithm 2. For all candidate poses, the vector from original target to candidate is checked against the original target orientation and the candidate only is used if the orientation difference is within a threshold. This is motivated by the fact that for observation of the target pose to be possible, it has to be in front of the robot. Simultaneously, this approach prevents the generation of target poses that lie behind walls. This could otherwise happen if thin walls are part of the environment. An example is shown in Fig 7.6.

The observation pose approach is used as a default setting when goal poses are sent to the robot. This notably includes cases where the robot tries to reach victim hypothesis for veri-fication autonomously, but also cases where the supervisor provides a pose goal. Using the described approach, carefree planning is possible for both higher level behavior components and supervisor. Either the requested pose can be reached; in that case, a plan to it is created. If it is not reachable, the alternative observation pose is computed and sent to the planner.

Data: Robot pose PV, goal pose Pg, map m, bounding box B Result: Collision-free refined goal pose Po

initialization;

Generate bounding box B around Pg; Initialize closest distance dmin to max float;

for all grid cells gi of m inside bounding box B do if gi not free cell then

Convert target pose from map to world frame;

Related documents