• No se han encontrado resultados

1. PLANTEAMIENTO DEL PROBLEMA

2.2 REFERENTES TEÓRICOS

2.2.1 Antecedentes de la conciliación en Colombia

4.3.1. Structure of the Hybrid Graph

Recall that the goal of our algorithm was to use the faster low-dimensional planning, except for areas of the environment where high-dimensional planning is necessary to ensure the feasibility of the resulting path and the desired cost sub-optimality bound. We want our hybrid state-space to capture this property—namely, we want GAD to consist largely of low-dimensional states, except for the areas where high-dimensional planning needs to be performed, represented by areas of high-dimensional states inGAD. To ensure path feasibil- ity in the high-dimensional regions ofGAD, we have to use high-dimensional transitions. In the low-dimensional areas we can use simpler low-dimensional transitions. However, recall that the transitions we have in THD andTLD connect two states of the same dimensional- ity, which do not allow us to transition from the low-dimensional to the high-dimensional regions. Therefore, we have to construct a transition set TAD that allows for transitions between states of different dimensionality.

4.3.2. Construction of the Hybrid Graph

Our algorithm iteratively constructs GAD, beginning with the low-dimensional state-space SLDand introducing a set of high-dimensional regionsRin it. We first explain how the high-

dimensional regions are being introduced intoGADand connected with the low-dimensional regions. The algorithm that decides when and where to introduce these regions will be explained later.

Once a high-dimensional regionr is introduced, the following changes are made toGAD. If a low-dimensional state XiLD falls inside a new high-dimensional region r ∈R, we replace it with its high-dimensional projection states inλ−1(XiLD). Thus,GAD contains both low-

dimensional and high-dimensional states. Notice that if a high-dimensional stateXHD is in SAD, then its low-dimensional projectionλ(XHD) is not in SAD, and also ifXHD 6∈SAD, then λ(XHD) ∈ SAD. Thus, for every state XHD in the original high-dimensional state- space, eitherXHD∈SADorλ(XHD)∈SAD (but not both). Adding new high-dimensional regions or increasing the sizes of existing regions requires the reconstruction of SAD and TAD, and thus, will produce a new instance ofGAD = (SAD, TAD).

Next we define the transition set TAD for the hybrid graph GAD as follows.

Definition 4.4 Transitions in GAD: For any state Xi ∈SAD:

• If Xi is high-dimensional (Xi ∈ SHD), then for all high-dimensional transitions

(Xi, XjHD) ∈ THD, if XjHD ∈ SAD then (Xi, XjHD) ∈ TAD. If XjHD 6∈ SAD,

then (Xi, λ(XjHD))∈TAD. That is, for high-dimensional states we allow only high-

dimensional transitions to other high-dimensional states if they fall inside SAD, or their low-dimensional projections (Fig. 5 lower left).

• If Xi is low-dimensional (Xi ∈ SLD), then for all low-dimensional transitions

(Xi, XjLD)∈TLD, ifXjLD ∈SAD then(Xi, XjLD)∈TAD and for all high-dimensional

transitions (X, XjHD)∈THD, where X ∈λ−1(Xi), if XjHD ∈SAD then(Xi, XjHD)∈

TAD. That is, for low-dimensional states we allow low-dimensional transitions if they lead to another low-dimensional state inSAD(Fig. 5 upper left), and high-dimensional transitions from their high-dimensional projections if they lead to a high-dimensional state inSAD (Fig. 5 right).

Notice, that the above definition of TAD allows for transitions between states of different

dimensionality. Figure 5 illustrates the set of transitions in the adaptive graph in the case of 3D (x, y, θ) path planning.

4.3.3. Mapping Hybrid Solutions to the High-Dimensional State-Space

Once we have computed a path through our hybrid graph GAD, which can contain low-

dimensional states and transitions, we need to be able to project it to the high-dimensional state-space in order to ensure that it is feasible and satisfies the desired solution cost sub- optimality bound. Therefore, we define a tunnel τ of radius w around a hybrid path πAD

as follows:

Definition 4.5 A tunnelτ of widthwaround a hybrid pathπADis a sub-graphτ = (Sτ, Tτ)

τ ⊆GHD such that Sτ ⊆SHD Tτ ⊆THD ∀XHD ∈SHD, XHD ∈Sτ iff ∃Xi ∈πAD s.t. dist(λ(XHD), Xi)≤w if Xi ∈SLD or dist(λ(XHD), λ(Xi))≤w if Xi ∈SHD ∀EHD= (Xi, Xj)∈THD, EHD ∈Tτ iff Xi ∈τ and Xj ∈τ

where dist is some pre-defined distance metric in SLD.

In other words,τ is a sub-graph ofGHD, and thus consists only of high-dimensional states and transitions. Moreover, τ contains all high-dimensional states XHD if they fall within distance w of some state Xi ∈ πAD. We include in τ all transitions (Xj, Xk) from THD

such that both Xj and Xk are in τ. It is important to note that the above definition of τ

for tunnel width w = 0 becomes equivalent to the sub-graph produced by projecting the hybrid path πAD to the high-dimensional state-space SHD through the projection function

λ−1. This λ−1 projection method can be used when no distance metric is available in the low-dimensional state-space. The above definition, however, allows for more flexibility when mapping hybrid paths into the high-dimensional state-spaceSHD. To produce a high-

(a)Initial 2D/3D path (b) Tunnel around path (c) Tracking in tunnel (d) Add HD region at point of failure

(e)2D/3D path (f) Tunnel around path (g)Tracking in tunnel (h) Add HD region at point of failure

(i)2D/3D path (j)Tunnel around path (k)Tracking in tunnel (l)Final trajectory

Figure 6: Example of the iterative process of Planning with Adaptive Dimensionality on simple map in the context of 3D (x,y,heading) path planning for a non-holonomic vehicle. Start: top left; goal: bottom right; light gray circles: 3D regions; darker gray outer circles: borders between 2D and 3D regions consisting of 2D states which have valid 3D transitions going into the 3D areas; white: 2D regions; black bars: obstacles.

dimensional path from a hybrid path πAD, we construct a tunnel τ around πAD; then we

perform a graph search from start to goal in τ, which is a small sub-graph of the original high-dimensional state-space. The search, if successful, produces a fully high-dimensional pathπHD corresponding to our hybrid pathπAD.