Eric Chen

EECS Department, University of California, Berkeley

Technical Report No. UCB/

December 1, 2025

The integration of robotics with cloud and edge computing presents significant opportunities to enhance scalability, efficiency, and collaboration across diverse robotic systems. However, current cloud robotics systems pose challenges across networking, reliability, compute resource management, and data management. This dissertation presents FogROS, a comprehensive cloud robotics platform that addresses these fundamental challenges through four key components. First, we introduce Secure Global Connectivity (SGC), which enables configuration-free connectivity between robots and cloud services through globally unique cryptographic identifiers. The message routing between identifiers is achieved through a hybrid global Distributed Hash Table (DHT) and a local Peer-to-Peer (P2P) routing system. SGC maintains connectivity even as robots roam across different networks while ensuring security. Second, we develop Probabilistic Latency Reliability (PLR), a framework that achieves reliable operation on commodity cloud infrastructure through multiple independent networks and compute resources. Our theoretical analysis leads to the LSC Impossibility Triangle theorem, which proves that providing replicated resources with uncorrelated failures can reduce failure probability exponentially. Based on the theoretical analysis, we design and implement different protocols, Anycast, FaultTolerance, and LatencyReliablity, depending on available independent resources. Third, we present FogROS2-Config, which automates cloud resource configuration and enables seamless integration of cloud resources into existing robot environments. The system includes intelligent resource selection across major cloud providers and supports both general-purpose computing and specialized hardware like GPUs. Fourth, we introduce RoboDM, an efficient cloud-based toolkit for collecting, sharing, and learning with robot data. The system streamlines storage for vision, language, and action data via a unified container format. Through extensive experimentation and real-world cloud robotics deployment, we demonstrate FogROS’s effectiveness across diverse applications including visual SLAM, grasp planning, and distributed fleet learning. The platform achieves up to 45x speedup in motion planning tasks compared to traditional approaches and 3.7x anomalous latency reduction. These results establish FogROS as a robust foundation for scalable, reliable, and efficient cloud robotics applications.

Advisors: John D. Kubiatowicz

---

BibTeX citation:

@phdthesis{Chen:31729,
    Author= {Chen, Eric},
    Title= {FogROS: A Cloud Robotics System},
    School= {EECS Department, University of California, Berkeley},
    Year= {2025},
    Number= {UCB/},
    Abstract= {The integration of robotics with cloud and edge computing presents significant opportunities to enhance scalability, efficiency, and collaboration across diverse robotic systems. However, current cloud robotics systems pose challenges across networking, reliability, compute resource management, and data management. This dissertation presents FogROS, a comprehensive cloud robotics platform that addresses these fundamental challenges through four key components. First, we introduce Secure Global Connectivity (SGC), which enables configuration-free connectivity between robots and cloud services through globally unique cryptographic identifiers. The message routing between identifiers is achieved through a hybrid global Distributed Hash Table (DHT) and a local Peer-to-Peer (P2P) routing system. SGC maintains connectivity even as robots roam across different networks while ensuring security. Second, we develop Probabilistic Latency Reliability (PLR), a framework that achieves reliable operation on commodity cloud infrastructure through multiple independent networks and compute resources. Our theoretical analysis leads to the LSC Impossibility Triangle theorem, which proves that providing replicated resources with uncorrelated failures can reduce failure probability exponentially. Based on the theoretical analysis, we design and implement different protocols, Anycast, FaultTolerance, and LatencyReliablity, depending on available independent resources. Third, we present FogROS2-Config, which automates cloud resource configuration and enables seamless integration of cloud resources into existing robot environments. The system includes intelligent resource selection across major cloud providers and supports both general-purpose computing and specialized hardware like GPUs. Fourth, we introduce RoboDM, an efficient cloud-based toolkit for collecting, sharing, and learning with robot data. The system streamlines storage for vision, language, and action data via a unified container format. Through extensive experimentation and real-world cloud robotics deployment, we demonstrate FogROS’s effectiveness across diverse applications including visual SLAM, grasp planning, and distributed fleet learning. The platform achieves up to 45x speedup in motion planning tasks compared to traditional approaches and 3.7x anomalous latency reduction. These results establish FogROS as a robust foundation for scalable, reliable, and efficient cloud robotics applications.},
}

EndNote citation:

%0 Thesis
%A Chen, Eric 
%T FogROS: A Cloud Robotics System
%I EECS Department, University of California, Berkeley
%D 2025
%8 December 1
%@ UCB/
%F Chen:31729