Real-World Deployment of a Hierarchical Uncertainty-Aware Collaborative Multiagent Planning System

TL;DR

This work tackles the challenge of deploying collaborative multiagent planning under real-world uncertainty by introducing a hierarchical planning system that couples abstract macro-actions on a navigational graph with groundable primitive actions on each robot. The central planner generates macro-actions that are executed by onboard bi-level planners, using robust execution strategies to handle disturbances and uncertain edge traversability. The approach is demonstrated in a real-world Jackal–Husky deployment, where collaborative planning reduces overall makespan by offloading sensing tasks to faster agents and waiting for critical information, despite occasional low-level failures and operator interventions. The work advances field robotics by showing that robust, multi-level planning can enable scalable collaboration under uncertainty and points to future work in automatic graph generation, GPS-denied grounding, and improved handling of communications in larger teams.

Abstract

We would like to enable a collaborative multiagent team to navigate at long length scales and under uncertainty in real-world environments. In practice, planning complexity scales with the number of agents in the team, with the length scale of the environment, and with environmental uncertainty. Enabling tractable planning requires developing abstract models that can represent complex, high-quality plans. However, such models often abstract away information needed to generate directly-executable plans for real-world agents in real-world environments, as planning in such detail, especially in the presence of real-world uncertainty, would be computationally intractable. In this paper, we describe the deployment of a planning system that used a hierarchy of planners to execute collaborative multiagent navigation tasks in real-world, unknown environments. By developing a planning system that was robust to failures at every level of the planning hierarchy, we enabled the team to complete collaborative navigation tasks, even in the presence of imperfect planning abstractions and real-world uncertainty. We deployed our approach on a Clearpath Husky-Jackal team navigating in a structured outdoor environment, and demonstrated that the system enabled the agents to successfully execute collaborative plans.

Figures (7)

• Figure 1: a) In this work, we discuss the deployment of a collaborative multiagent planner for a two agent team consisting of a Clearpath Jackal and a Clearpath Husky in a real-world structured outdoor environment. b) Consider the Jackal and Husky navigating from starts SJ and SH to goals GJ and GH, respectively, given a navigation graph (green), where one edge has unknown traversability (red). c) When the agents plan independently, the Husky attempts to navigate to the goal via the unknown edge, senses that the edge is untraversable, and then backtracks to the goal via a long, traversable path (trajectory shown in yellow-orange); meanwhile, the Jackal navigates directly to its goal (trajectory shown in blue-green). d) When the agents plan to collaboratively minimize team makespan, the quicker Jackal diverts from the shortest path to its goal to sense the traversability of the unknown edge for the Husky, while the Husky waits in place for additional information. After sensing that the edge is untraversable, both the Jackal and the Husky navigate to their respective goals.
• Figure 2: Overview of of the hierarchical planning system. Multiagent planning, which occurred at a centralized base station, generated macro-actions for each agent in the team. The macro-actions were then sent to the individual agents, where they were processed and executed as sequences of primitive actions, using information from each agent's individual core autonomy pipeline.
• Figure 3: An example collaborative team plan. The Jackal and Husky were tasked with navigating from SJ and SH to GJ and GH, respectively, by executing actions on a pre-defined abstract motion graph (green) where some edges had unknown traversabilities (red). In collaborative plans, agents took actions that were individually suboptimal to reduce the expected team makespan. Here, the Jackal diverted from its optimal path to the goal to sense and share the traversability of the unknown edge (a), while the Husky waited for the edge traversability information. Then, the Husky used the information to select the best traversable path to its goal, which reduced its expected plan cost (b-c).
• Figure 4: An example macro-action (orange) that represents an agent first navigating from node A to B to C, and then sensing the traversability of the unknown edge C-D (gold star). The inlay depicts goal adjustment, which enables an agent to navigate to a location close to the original GPS location associated with the target node, if navigation to the target GPS location is deemed infeasible. This adjustment enables the agent to be robust to various types of uncertainty during planning, including minor pose drift, minor map noise, and small changes in the environment structure.
• Figure 5: Examples of macro-action interrupts (received when the agents are at the red Xs) and their impacts on planning. When an agent received an interrupt message during primitive action execution, it completed its current primitive navigation action and any directly succeeding sensing actions, then terminated the current macro-action, communicated with the base station, and waited to receive an updated plan.