FarmBrain

NSF CISE VINES (Track 1) · Award 2548960

FarmBrain: Enabling Real-Time Agricultural Robotics with AI-Native Cellular Access Edge

Synopsis

Agricultural robotics is a critical answer to farm labor shortages and the need for more efficient food production, yet today’s systems face a fundamental economic barrier. Robots must be lightweight enough to navigate narrow farm environments — corn rows as narrow as 15 inches, strawberry tunnels with 12–15 inches of clearance — and at the same time powerful enough to run AI-driven perception and planning within sub-100-millisecond real-time constraints. The result is robots that cost hundreds of thousands of dollars each, an economics that prevents widespread deployment.

FarmBrain is an AI-native 5G edge architecture that resolves this impasse. Lightweight robots access sophisticated AI capabilities through low-latency private 5G connectivity to nearby on-farm edge servers. High-frequency reactive control loops (motion control, safety) stay on the robot, while low-frequency deliberative loops (3D scene reconstruction, vision-language inference, motion planning) run as shared services on the edge. Sharing expensive compute across a fleet of simple robots changes the economics of agricultural robotics while preserving the real-time responsiveness and safety required for delicate manipulation.

Today: self-contained robot
Today: a robot carries a high-performance on-board computer that runs both the reactive and the deliberative control loop, making it heavy, expensive, and power-hungry. High-performance on-board computer GPU-class · big battery · cooling Reactive loop 100–1000 Hz · motion, safety Deliberative loop 3D scene, VLMs, motion planning Strawberry-picking robot Heavy · Expensive · Power-hungry
FarmBrain: split, multi-rate control
FarmBrain: the robot keeps only the reactive loop on a low-power embedded processor; the deliberative loop runs on a 5G edge server shared by many robots. Plans flow down to the robot and sensor data flows up over a farm-wide private 5G network. Embedded processor Reactive loop 100–1000 Hz Farm-wide private 5G network plans sensors 5G edge server shared GPUs Deliberative loop AI perception Lightweight robot fleet shares one edge Lightweight · Inexpensive · Low-power
Today's robots run both control loops on a heavy on-board computer. FarmBrain keeps only the high-frequency reactive loop on the robot and runs the deliberative loop on a 5G edge server that the whole fleet shares.

Challenges

Programming split, multi-rate control

Robotics frameworks lack abstractions for cooperating control loops with explicit SLOs and for cross-network shared state with non-blocking, bounded-staleness guarantees under wireless jitter.

Tight end-to-end SLOs under variability

Wireless links add stochastic delay, and uplink and GPU contention inflate latency. The RAN cannot see application deadlines, and the edge cannot see network delay.

End-to-end resilience

Robots have physical state that cannot simply be "restarted." Local and remote failures must both trigger time-bounded safe holds, and RAN failures need sub-millisecond failover.

Manageability for non-experts

Observability for robots, RAN, and edge is siloed and time-misaligned, and root causes span non-linear, time-lagged dependencies across layers. Farm operators need actionable guidance.

Research thrusts

FarmBrain architecture. Thrust 1, FarmLang, is the programming model. Thrust 2, the compiler and runtime, deploys programs across an on-farm edge server and farm robots connected by a farm-wide private 5G network. Thrust 3, AI-assisted management, collects telemetry from all of them, diagnoses problems, and guides the farm operator. FarmLang Programming model THRUST 1 Compiler & runtime Partition, schedule THRUST 2 deploy On-farm edge server Farm-wide private 5G network Farm robots telemetry Management Diagnose, recover THRUST 3 guidance Farm operator
FarmBrain architecture. Developers write FarmLang programs (Thrust 1); the compiler and runtime deploy them across farm robots and a shared on-farm edge server connected by a farm-wide private 5G network (Thrust 2); AI-assisted management diagnoses problems from cross-layer telemetry and keeps the system running for non-expert operators (Thrust 3).

A Programming Model for Multi-Rate Control Loops with SLOs

FarmLang lets robotics developers describe a task as a set of cooperating control loops, each with its own timing requirement. They write the task once; coordinating it across the robot and the edge over an unreliable wireless link is left to the system. FarmLang also gives developers one simple way to say how the robot should stay safe when something goes wrong, whether the failure is physical (a slipping gripper) or in the network (a dropped connection).

An SLO- and Resilience-Aware Compiler and Runtime for the AI-Native Edge

The FarmBrain compiler automatically splits each program between the robot and the edge. Time-critical control stays on the robot, and heavy AI tasks move to shared edge servers. At run time, FarmBrain scales edge services up and down as the workload changes, recovers quickly from failures, and schedules both the 5G network and the edge GPUs so that every robot meets its deadlines.

AI-Assisted Management and Resilience for Complex Infrastructure

FarmBrain monitors itself across robots, the 5G network, and the edge. It uses AI to find the root cause when performance degrades, then fixes the problem automatically or tells farm operators what to do in plain language. If 5G infrastructure fails, it fails over within milliseconds while the robots hold a safe state.

Prototype and evaluation

FarmBrain will be prototyped on a private 5G edge testbed at UT Austin together with an agricultural robot platform built on a mobile rover base and robot arm. It will be validated on fruit-picking, navigation, and crop-inspection tasks with multiple robots. A controlled physical “twin” testbed with artificial fruit allows evaluation year-round, followed by deployment on real farms with industry partners.

Broader impacts

If successful, FarmBrain will lower the barrier to deploying advanced agricultural robotics, making AI-driven robots economically viable for small and medium-scale farms as well as large ones. The AI-native edge platform offers a replicable model for real-time edge computing in other safety-critical domains such as autonomous vehicles, industrial automation, and emergency response. The project will release the FarmBrain framework, including software, programming tools, example applications, and benchmarks, as open source. Research findings will be integrated into computer systems and robotics curricula at UT Austin, including the Freshman Research Initiative on Agricultural Robotics. Results will also be shared with industry partners and at the LDOS Research Symposium.

Open-source software, data, and results will be made available on this website.