Multi-agent path topology in support of socially competent navigation planning

Abstract

We present a navigation planning framework for dynamic, multi-agent environments, where no explicit communication takes place among agents. Inspired by the collaborative nature of human navigation, our approach encodes the concept of coordination into an agent's decision making through an inference mechanism about collaborative strategies of collision avoidance. Each such strategy represents a distinct avoidance protocol, prescribing a distinct class of navigation behaviors to agents. We model such classes as equivalence classes of multi-agent path topology, using the formalism of topological braids. This formalism may naturally encode any arbitrarily complex, spatiotemporal, multi-agent behavior, in any environment with any number of agents into a compact representation of dual algebraic and geometric nature. This enables us to construct a probabilistic inference mechanism that predicts the collective strategy of avoidance among multiple agents, based on observation of agents' past behaviors. We incorporate this mechanism into an online planner that enables an agent to understand a multi-agent scene and determine an action that not only contributes progress towards its destination, but also reduction of the uncertainty of other agents regarding the agent's role in the emerging strategy of avoidance. This is achieved by picking actions that compromise between energy efficiency and compliance with everyone's inferred avoidance intentions. We evaluate our approach by comparing against a greedy baseline that only maximizes individual efficiency. Simulation results of statistical significance demonstrate that our planner results in a faster uncertainty decrease that facilitates the decision-making process of co-present agents. The algorithm's performance highlights the importance of topological reasoning in decentralized, multi-agent planning and appears promising for real-world applications in crowded human environments.

Where to go Next: Learning a Subgoal Recommendation Policy for Navigation in Dynamic Environments

“…1) Navigation Among Crowds: Past work on navigation in cluttered environments often focuses on interaction models using geometry [8], [9], physics [10], topologies [11], [12], handcrafted functions [13], and cost functions [14], [14] or joint probability distributions [15] learned from data. While accurate interaction models are critical for collision avoidance, this work emphasizes that the robot's performance (time-togoal) is highly dependent on the quality of its cost-to-go model (i.e., the module that recommends a subgoal for the local planner).…”

Algorithms for Microscopic Crowd Simulation: Advancements in the 2010s

“…Their method defines a topological strategy as a set of decisions to move around agents/obstacles along the left or right. (As discussed in Section 8.2, Mavrogiannis et al [MK19] explored similar ideas for collision avoidance among a small number of agents.) Such a strategy can be computed for the agent's global path, for its preferred velocity that results from path following, and for the velocity that results from collision avoidance.…”

Social Trajectory Planning for Urban Autonomous Surface Vessels

“…The work of Knepper and Rus [16] proposes a collision avoidance method that, based on the idea of equivalence relation on local paths [17], encodes the avoidance behavior inherent to pedestrians. In addition, using the formalism of topological braids, recent work [18] proposes a framework that is capable of generating legible collision-free paths for multirobot navigation.…”

“…, θ 5 ) using trajectory data from the AIS dataset. 5 We adopt the "softened" value iteration, as in [13], to compute a value function for (18) and using it, we find a (deterministic) policy μ that derives vessel trajectories that most likely occur under the MaxEntIRL approach.…”