Congestion-Aware Policy Synthesis for Multirobot Systems

Abstract

Multi-robot systems must be able to maintain performance when robots get delayed during execution. For mobile robots, one source of delays is congestion. Congestion occurs when robots deployed in shared physical spaces interact, as robots present in the same area simultaneously must manoeuvre to avoid each other. Congestion can adversely affect navigation performance, and increase the duration of navigation actions. In this paper, we present a multi-robot planning framework which utilises learnt probabilistic models of how congestion affects navigation duration. Central to our framework is a probabilistic reservation table which summarises robot plans, capturing the effects of congestion. To plan, we solve a sequence of single-robot time-varying Markov automata, where transition probabilities and rates are obtained from the probabilistic reservation table. We also present an iterative model refinement procedure for accurately predicting execution-time robot performance. We evaluate our framework with extensive experiments on synthetic data and simulated robot behaviour.

Decision-making under uncertainty for multi-robot systems

“…However, the model of each robot is extremely impoverished, with all robots being represented as tokens which are indistinguishable. Thus, in [34], we proposed a methodology that models MRS using a joint state, composed of the Cartesian product of each robot's local state features and a set of global state features, but still allows the modelling of asynchronous execution in continuous time. We achieve this by extending MAs to the multi-robot case.…”

“…Contexts allow us to distinguish between different situations when executing an action, without considering the full joint state. This enables the construction of MRMA models from empirical data, and in [34] we propose data gathering policies. Finally, we introduce the context-aware multi-agent simulator (CAMAS), a discrete event simulator based on the MRMA model.…”

“…The MRMA model presented in [34] is an accurate representation of the global state of the MRS and can be used to evaluate both global and local properties of a set of policies. However, it is not suitable for planning due to its joint state representation.…”

“…To plan using context-aware models, we have taken a more local perspective where each robot has its own goal and the team objective is to find a makespanminimising solution (i.e. a solution that minimises the cost for all robots to achieve their goals) [33,35]. The congestion-aware planner (CAP) focuses on congestion as a form of context and the impact of this context on navigation duration.…”

“…We believe there are two main challenges related to this (i) the asynchronous action execution of teams of mobile robots is not considered in state-of-the-art MARL approaches; and (ii) there is a lack of simulators of multi-robot behaviour that can quickly and accurately provide large amounts of data for training. We believe CAMAS [34] is appropriate to tackle point (ii), and are currently researching adaptations of MARL algorithms to consider asynchronous action execution, intending to develop a MARL algorithm that enables context-aware multi-robot coordination. The use of MARL approaches will allow us to scale up to larger problems and relax observability and communication assumptions during execution, since robots will condition their actions only on (local) contexts.…”