OpenBot-Fleet: A System for Collective Learning with Real Robots

TL;DR

OpenBot-Fleet tackles the problem of scalable, continual policy improvement for real-robot navigation by uniting a low-cost, smartphone-enabled fleet with a cloud-based learning backend. The approach combines self-supervised perception learning, simulation-pretrained control, and online reinforcement learning with an asynchronous replay buffer deployed via TensorFlow Lite on smartphones, enabling data collection from a large fleet and rapid policy updates. Key contributions include an open-source, end-to-end cloud robotics stack, a large real-world crowd-sourced dataset, and demonstrated zero-shot and few-shot generalization to unseen homes with strong performance, illustrating practical impact for scalable autonomous navigation in the wild. The work advances cloud robotics by showing that continual learning across thousands of real robots is feasible at a fraction of the cost of industrial deployments, with potential for rapid deployment in varied real-world settings.

Abstract

We introduce OpenBot-Fleet, a comprehensive open-source cloud robotics system for navigation. OpenBot-Fleet uses smartphones for sensing, local compute and communication, Google Firebase for secure cloud storage and off-board compute, and a robust yet low-cost wheeled robot toact in real-world environments. The robots collect task data and upload it to the cloud where navigation policies can be learned either offline or online and can then be sent back to the robot fleet. In our experiments we distribute 72 robots to a crowd of workers who operate them in homes, and show that OpenBot-Fleet can learn robust navigation policies that generalize to unseen homes with >80% success rate. OpenBot-Fleet represents a significant step forward in cloud robotics, making it possible to deploy large continually learning robot fleets in a cost-effective and scalable manner. All materials can be found at https://www.openbot.org. A video is available at https://youtu.be/wiv2oaDgDi8

Figures (16)

• Figure 1: Overview of OpenBot-Fleet, a comprehensive cloud robotics system for navigation which supports both simulated agents and real robots. Our wheeled robots are robust, low-cost, and use a smartphone for sensing and compute. A smartphone app allows for teleoperation and control by a navigation policy, and secure communication with the cloud system to upload robot experience and receive policies. We train a self-supervised perception backbone using data from 72 robots and pretrain a controller in simulation. The full navigation policy is evaluated in unseen real-world environments.
• Figure 2: Our robot with a connected smartphone, adapted from OpenBot mueller2021openbot for shipping to and use by crowd workers. In contrast to the original DIY version of OpenBot, it has a robust injection-molded plastic shell, integrated electronics (visible in the bottom-right image), and improved cable routing.
• Figure 3: Features of our app which runs on the smartphone connected to each robot. It communicates with the robot via a two-way serial link to send controls to it (e.g., while executing a control policy) and to receive sensor data for logging. A Bluetooth game controller can be used for teleoperation.
• Figure 4: Unseen real-world environments used for evaluation.
• Figure 5: Overview of our system for collective learning with a fleet of mobile robots. The cloud backend (B) publishes recording requests that are observed by the robot fleet (A) via real-time listeners. After the successful upload of a recording, the backend either stores the data in a database for offline learning or notifies an online learning process to consume the data directly. When a new policy is available, it is automatically redeployed to the robot fleet. To validate our system, we use a large fleet of mobile robots to train a perception backbone using self-supervised learning and also implement goal-conditioned imitation learning codevilla2018end based on PilotNet bojarski2016end as a baseline (D). We have also built a fast procedural simulator (ProceduralSim) based on Unity ML Agents juliani2018unity for pre-training navigation policies and integrated the photorealistic simulator SpearSim spear to perform extensive evaluations before real-world deployment (E). We pre-train a control policy in Unity based on privileged perception information using reinforcement learning. Both the perception backbone and pre-trained control policy are packaged into TensorFlow-lite models for deployment to the smartphones, which execute them to control the robots (A). Once deployed, the system can be further finetuned via online learning. Details are given in Section III of the main paper.