SwiftBot: A Decentralized Platform for LLM-Powered Federated Robotic Task Execution
A new system uses LLMs and peer-to-peer networking to coordinate robot swarms, cutting task latency by up to 5.4x.
A research team led by YueMing Zhang and five others has unveiled SwiftBot, a novel decentralized platform designed to coordinate fleets of robots using large language models (LLMs) and peer-to-peer networking. Accepted by IEEE CCGrid 2026, the system directly addresses key limitations in current robotic coordination: rigid, hand-coded planners, reliance on a central control server, and inefficient static resource allocation. SwiftBot's core innovation is its co-design of semantic understanding and federated resource management, allowing robots to understand complex natural language commands and then dynamically share computational workloads across a decentralized network.
At its heart, SwiftBot uses an LLM to decompose high-level instructions (like 'survey the building and report damage') into executable sub-tasks. These sub-tasks are then intelligently assigned and executed across a swarm of heterogeneous robots connected via a Distributed Hash Table (DHT) overlay—a decentralized network that eliminates the single point of failure of a central server. The platform's intelligent container orchestration enables 'federated warm container migration,' meaning pre-loaded software environments can be shared between robots to drastically reduce startup times as workloads shift.
Benchmark results are significant. SwiftBot achieves 94.3% accuracy in task decomposition across diverse scenarios. More importantly, it delivers substantial performance gains: reducing task startup latency by 1.5 to 5.4 times, cutting average training latency by 1.4 to 2.5 times, and improving system responsiveness under high load (tail latency) by 1.2 to 4.7 times. This demonstrates that a decentralized approach, when combined with advanced AI for planning, can surpass the efficiency of traditional centralized systems, especially as the number of robots scales.
- Achieves 94.3% accuracy in using LLMs to decompose natural language commands into executable robot sub-tasks.
- Reduces robotic task startup latency by 1.5-5.4x using federated warm container migration across a peer-to-peer network.
- Eliminates the need for a central server by using a DHT overlay, enabling scalable, fault-tolerant collaboration for robot swarms.
Why It Matters
Enables scalable, resilient coordination for robot fleets in search & rescue, logistics, and manufacturing without a central point of failure.