The Cambridge RoboMaster: An Agile Multi-Robot Research Platform

TL;DR

A tightly integrated hardware, control, and simulation software stack on a fleet of holonomic ground robot platforms designed with this motivation, presenting an in-depth review of other platforms currently available and new experimental validation of the system's capabilities.

Abstract

Compact robotic platforms with powerful compute and actuation capabilities are key enablers for practical, real-world deployments of multi-agent research. This article introduces a tightly integrated hardware, control, and simulation software stack on a fleet of holonomic ground robot platforms designed with this motivation. Our robots, a fleet of customised DJI Robomaster S1 vehicles, offer a balance between small robots that do not possess sufficient compute or actuation capabilities and larger robots that are unsuitable for indoor multi-robot tests. They run a modular ROS2-based optimal estimation and control stack for full onboard autonomy, contain ad-hoc peer-to-peer communication infrastructure, and can zero-shot run multi-agent reinforcement learning (MARL) policies trained in our vectorized multi-agent simulation framework. We present an in-depth review of other platforms currently available, showcase new experimental validation of our system's capabilities, and introduce case studies that highlight the versatility and reliability of our system as a testbed for a wide range of research demonstrations. Our system as well as supplementary material is available online. https://proroklab.github.io/cambridge-robomaster

Figures (9)

• Figure 1: Shown from left to right: 1) A close-up of one RoboMaster equipped with an NVidia Jetson Orin NX and a forward-facing camera 2) Two RoboMasters in a SLAM-based collision avoidance scenario 3) Five RoboMasters moving through a narrow constriction after breaking formation blumenkamp_framework_2022 4) Eight RoboMasters in a multi-robot navigation scenario trained in VMAS bettini_vmas_2022.
• Figure 2: We compare a variety of different research and commercial robotic platforms concerning their tradeoff between cost per unit and maximum velocity. We differentiate between Ackermann, Differential, Airborne and Omnidirectional platforms as well as between platforms specifically targetting multi-agent research. A clear correlation can be seen between speed and cost, as well as a trend for multi-agent platforms to be low-cost, but as a consequence slow. We draw the Pareto front as a dashed line, with our proposed platform pushing the limits on the cost and speed tradeoff.
• Figure 3: We modify the RoboMaster S1 Platform by removing the USB Camera, Smart Controller, Blaster and Turret and by replacing them with either a Jetson Orin on an AverMedia D131 carrier board or a Raspberry Pi. The platform is equipped with two WiFi modules, one for infrastructure communication (dark blue) and one specifically chosen for ad-hoc peer-to-peer communication (green).
• Figure 4: The CAN-based protocol used in the RoboMasters. DJI uses CAN as transport layer to allow multiple connected devices on the bus to broadcast information. Due to the limited length of CAN frames, a serial transport protocol is built on top, with a higher-level package definition permitting publisher/subscriber-type communication architectures.
• Figure 5: We utilize two wireless networks, one for infrastructure (blue) and one for ad-hoc peer-to-peer communication between agents. The infrastructure network is used to communicate with fixed infrastructure. The multi-agent user interface to control multiple agents efficiently is displayed in the top. The ad-hoc network is used for experiments requiring decentralized communication.