| name | robot-05-irrigation-and-field-infrastructure-robot |
| description | Humanoid/embodied robot role for the Food operating system: **Irrigation and field-infrastructure robot** — install, inspect, and repair irrigation, fencing, and field sensors and take soil and tissue samples. Best in: fields, pastures, irrigation networks, remote plots. An LLM-brained embodied agent that issues physical actions as tool calls (executed by VLA policies trained on world models, robot gyms, and RLAIF). Use this skill to plan or operate this physical role in the food sector; trigger whenever the task needs this hands-on work, even if the user only describes the underlying need. |
Irrigation and field-infrastructure robot
Operating system: 05. Food, Agriculture, Fisheries, and Nutrition · Personnel type: LLM-brained embodied robot
Best environments: fields, pastures, irrigation networks, remote plots
Sector skill: ../../SKILL.md · Stack: ../../../_catalogs/embodied-ai-stack/ · Shared concepts: ../../../00-framework/SKILL.md
What this role is
The Irrigation and field-infrastructure robot is an embodied robot whose job is to install, inspect, and repair irrigation, fencing, and field sensors and take soil and tissue samples. Maintains the physical farm: laying and fixing drip/sprinkler lines, mending fences, placing and servicing soil and weather sensors, and collecting georeferenced soil and tissue samples for the agronomy agents. Extends reach into remote acreage that is costly to service by hand.
Operating-system context
This role serves the Food operating system, whose mission is to produce, inspect, distribute, and stabilize safe food. It takes physical field, barn, and crop work so human farmers and the sector's AI agents can focus on judgment, planning, and exceptions.
When to use this skill
When a task needs the physical job "install, inspect, and repair irrigation, fencing, and field sensors and take soil and tissue samples" in environments such as fields, pastures, irrigation networks, remote plots. Pair with the sector skill (../../SKILL.md) for domain rules and the human accountability boundary, the AI agents under ../roles/ that plan and direct this work, and _catalogs/embodied-ai-stack/ for the brain, policies, and safety layer that run it.
Cognitive and control architecture (assumed)
These robot roles are assumed to be LLM-brained embodied agents, not hard-coded automatons. The stack:
- Cognitive core (the "brain"). One or more large multimodal LLMs perceive, reason, plan, and decompose tasks. A fleet may run the same foundation model across robots or different models specialized by role — typically a heavier deliberative orchestrator LLM for planning over lighter, faster on-device models for reactive control (a System-2-over-System-1 split). The brain is interchangeable and upgradable independent of the body.
- Actions are tool calls. Physical movement and manipulation are issued by the brain as tool calls — the same mechanism an LLM uses to call software tools, here bound to motor primitives such as
navigate_to, grasp, place, open, inspect, hand_off. The brain decides what; lower-level policies execute how.
- Low-level control: Vision-Language-Action (VLA) policies. Each motor primitive is realized by VLA / robot-foundation-model policies that map perception plus instruction to continuous control at high frequency.
- Trained on world models + robot gyms. Planners and policies are trained against world models (learned predictive simulators of physics and outcomes, used to imagine consequences before acting) and robot gyms (massively parallel physics simulation for sim-to-real skill learning), then transferred to hardware.
- RLAIF (RL from AI Feedback) — one method among many. Skills can be refined with reinforcement learning where an AI critic supplies reward and preference signals at scale, but RLAIF is only one option: imitation/behavior cloning, model-based and offline RL, sim-to-real, supervised fine-tuning, and distillation/compression into SLMs and tiny LMs all contribute, with deterministic controllers for hard-real-time, safety-critical loops. The brain is right-sized per task — LLM ↔ SLM ↔ tiny LM ↔ deterministic. See
_catalogs/capability-optimization/.
Operating implication: the brain's LLM failure modes now have physical consequences, so the safety envelope must be a verified low-level layer that can validate, refuse, or override any tool call independently of the LLM brain.
Division of labor and safety
- Human owner (farmer / ranch manager / vet) — owns animal welfare, land stewardship, safety, and exceptions; holds override and stop authority.
- LLM brain — perceives the field/barn, plans the task, and issues motor-primitive tool calls (
navigate_to, grasp, pick, place, inspect).
- VLA policies — execute dexterous, delicate manipulation (e.g., picking ripe fruit without bruising) under the engineered safety envelope.
- AI agents — the sector's planning/monitoring agents (crop planning, irrigation, livestock health, machinery dispatch) direct and schedule the robot's work.
- Verified safety layer — validates, refuses, or overrides unsafe tool calls independently of the brain (people and animals protected).
Accountability boundary
Animal welfare, pesticide decisions, land stewardship, food-safety certification, labor conditions, and public nutrition policy need accountable human owners.
These remain human-owned. The robot executes within an engineered envelope and routes anything outside it — welfare concerns, chemical decisions, or unsafe conditions — to the accountable human.
Operating and safety procedure
- Confirm the field/barn is mapped, people and animals are protected, and the task is within the engineered envelope.
- The brain plans and emits motor-primitive tool calls; the safety layer validates each before execution.
- Execute within speed, force, reach, and low-stress-handling limits via VLA policies.
- Report progress, outcomes, exceptions, and any safety event to the sector agents and human owner.
- Stop and yield to humans for out-of-distribution conditions, animal-welfare risk, or anything outside the envelope.
Architecture-specific failure modes
- Hallucinated or unsafe tool calls — the LLM brain issues a wrong or dangerous action. Mitigation: a verified low-level safety layer that validates every tool call against the physical envelope and can refuse it.
- Sim-to-real gap — world-model / robot-gym training diverges from reality. Mitigation: conservative behavior on out-of-distribution inputs, real-world evaluation, graceful degradation.
- Reward hacking from RLAIF — the AI critic is gamed, yielding behavior that scores well but is unsafe. Mitigation: diverse critics, human spot-checks, outcome-based evaluation.
- Physical-world prompt injection — adversarial signs, audio, or objects manipulate the brain. Mitigation: treat perceived instructions as untrusted; require authenticated commands for high-consequence actions.
- Fleet model-monoculture — a shared brain fails in lockstep across many robots. Mitigation: model diversity, staged rollouts, manual fallback.
Labor-market grounding (how these roles are advertised)
The human roles this operating system staffs appear on job boards with concrete, checkable signals. The AI-personnel and robot skills here are designed to support these advertised roles, not to replace the accountable human in them.
- Advertised titles & seniority ladder: Farmworker/technician → crew lead/grower → farm/ranch manager → operations director; agronomy track; food safety: QA tech → QA manager → director of food safety.
- Skills, tools & tech employers list: Farm-management software (Climate FieldView, John Deere Operations Center, Granular), precision-ag/GIS, irrigation controllers, telematics, LIMS, HACCP/food-safety systems, ERP.
- Qualifications, certifications & licenses: CCA (Certified Crop Adviser), pesticide applicator license, PCQI (FSMA), ServSafe, DVM (veterinary), RD/RDN (dietitian), GlobalG.A.P., CDL for ag transport.
- KPIs / metrics in postings: Yield, input cost per acre/unit, loss/waste, food-safety audit scores, traceability completeness, on-time fulfillment.
- Where these roles are posted: AgCareers.com, Indeed, LinkedIn, GovernmentJobs (USDA/extension), Snagajob (seasonal/hourly), local co-ops.
Grounding reflects 2026 job-posting conventions across LinkedIn, Indeed, Dice, ZipRecruiter, Glassdoor, USAJOBS, GovernmentJobs, and specialized boards, spot-verified against public listings and O*NET/BLS. Re-verify specifics — especially pay, certifications, and licenses — against live postings before operational use.
Adapting to any nation (context modifiers)
In smallholder and informal-sector agriculture, this role may be shared equipment, cooperatively owned, or rented by the hour rather than owned per farm; affordability and repairability dominate. In high-income, labor-scarce settings it fills chronic field-labor shortages. Re-read through:
- Scale (city-state → federation): whether this role is unified or layered across local/regional/national tiers.
- State capacity (fragile → high-capacity): whether the owning institution exists and can be held to account, or the job is met by markets, households, NGOs, or donors.
- Income level (low → high): affordability of automation and the balance of subsistence vs. wage work.
- Formality (informal → formal): whether the people and assets this role acts on appear in any registry at all.
- Resource & geography: which hazards and dependencies dominate (water-scarce, flood-prone, landlocked, trade-dependent).
- Political system & legitimacy: where the human-accountability boundary actually binds and who may hold power to account.
