RobotFleet: An Open-Source Framework for Centralized Multi-Robot Task Planning

TL;DR

RobotFleet addresses the challenge of coordinating heterogeneous robot fleets to achieve multiple goals by combining LLM-driven task planning, centralized task allocation, and containerized execution. It builds dependency DAGs from high-level goals, supports multiple planning strategies (Per-Goal DAG, Big-DAG, Monolithic Prompt) and two allocation methods (LLM-based and MILP), with a replanning loop that updates the world state and reallocates tasks. The framework emphasizes modularity with abstract base classes, a CLI for robot registration, and Docker-based deployment to enable scalable real-world deployments. Real-robot demonstrations and scalability analyses validate the design, showing effective planning and replanning across a Toyota HSR and a LoCoBot, while highlighting trade-offs between planning structure and allocation strategy.

Abstract

Coordinating heterogeneous robot fleets to achieve multiple goals is challenging in multi-robot systems. We introduce an open-source and extensible framework for centralized multi-robot task planning and scheduling that leverages LLMs to enable fleets of heterogeneous robots to accomplish multiple tasks. RobotFleet provides abstractions for planning, scheduling, and execution across robots deployed as containerized services to simplify fleet scaling and management. The framework maintains a shared declarative world state and two-way communication for task execution and replanning. By modularizing each layer of the autonomy stack and using LLMs for open-world reasoning, RobotFleet lowers the barrier to building scalable multi-robot systems. The code can be found here: https://github.com/therohangupta/robot-fleet.

Figures (4)

• Figure 1: RobotFleet is an open-source framework for coordinating multiple goals across a fleet of heterogeneous robots by leveraging LLMs for task planning and allocation.
• Figure 2: Overview. RobotFleet consists of three phases: task planning from goals, task allocation to robots, and task execution on optionally-containerized robot workers. For task planning, we support multiple types of task planners which generate plans in the form of DAGs. These tasks in the generated plans are then allocated to and executed across multiple robots.
• Figure 3: We show the end-to-end real-robot execution from RobotFleet based on two provided high-level goals, which were decomposed into dependent subtasks, as shown in the DAG in Figure \ref{['fig:fig4']}. We use a Locobot navigation robot and an HSR mobile manipulator and are able to effectively separate subtasks to accomplish multiple goals.
• Figure 4: The directed acyclic graph (DAG) representing the task plan and their dependencies created by RobotFleet given the goals described in Figure \ref{['fig:fig3']}. We find that the generated DAG can effectively decompose multiple general goals into dependent multi-robot subtasks.