Collision avoidance for aerial vehicles in multi-agent scenarios

Abstract

This article describes an investigation of local motion planning, or collision avoidance, for a set of decision-making agents navigating in 3D space. The method is applicable to agents which are heterogeneous in size, dynamics and aggressiveness. It builds on the concept of velocity obstacles (VO), which characterizes the set of trajectories that lead to a collision between interacting agents. Motion continuity constraints are satisfied by using a trajectory tracking controller and constraining the set of available local trajectories in an optimization. Collision-free motion is obtained by selecting a feasible trajectory from the VO’s complement, where reciprocity can also be encoded. Three algorithms for local motion planning are presented—(1) a centralized convex optimization in which a joint quadratic cost function is minimized subject to linear and quadratic constraints, (2) a distributed convex optimization derived from (1), and (3) a centralized non-convex optimization with binary variables in which the global optimum can be found, albeit at higher computational cost. A complete system integration is described and results are presented in experiments with up to four physical quadrotors flying in close proximity, and in experiments with two quadrotors avoiding a human.

Appendices

Appendix 1: Extension to homogenous group of agents

This appendix describes an extension of the local motion planning method of Sect. 4 towards an homogeneous group of agents (having the same control parameters).

With the assumption that all agents have the same control parameters ($\omega _0, \omega _1$) for the local trajectories (Sec. 7.4), the $\varepsilon$ enlargement of the agents is not required. This is achieved by substituting the $VO$ constraint (Constraint 3 of Sec. 4.5) by a 3D extension of the control obstacle $C^M-CO$ introduced by Rufli et al. (2013). The $C^M-CO$ characterizes the ($n$-differentiable) control trajectories in collision and is computed by formulating in relative candidate reference velocity space the full trajectories (Eq. (22) for $M=2$). Linearization of the constraint is still required and is done with respect to the current velocity. The algorithms described in this paper can be applied thereafter.

Relying on the concept of differential flatness for a quadrotor vehicle (Mellinger and Kumar 2011), if $M=5$ is used, the quadrator would, in theory, be able to perfectly track the control trajectory. In this case the full state (up to the fifth derivative) shall be known for all agents.

Appendix 2: Equations repulsive velocity

For the repulsive velocity field of Fig. 4, left, the repulsive velocity for agent $i\in {\mathcal {A}}$ is given by

where $V_r$ is the maximum repulsive force and $D_r$, $D_h$ the preferred minimal inter-agent distance in the X–Y plane and in the Z component respectively.

Appendix 3: Equations linearization of VO

Denote $\bar{h}_{ij} = \bar{h}_i + \bar{h}_j$ and $\bar{r}_{ij} = \bar{r}_i + \bar{r}_j$.

The non-convex constraint ${\mathbb {R}}^3 \setminus {\mathcal {VO}}_{ij}^{\tau }$ is linearized to obtain a convex problem. For an approximation with five linear constraints, as in Fig. 7, the linear constraints are given by

where $H_{ij}^1$ and $H_{ij}^2$ represent avoidance to the right / left, $H_{ij}^3$ and $H_{ij}^4$ above / below and $H_{ij}^5$ represents a head-on maneuver, which remains collision-free up to $t=\tau$.