What makes field robotics hard

Field robotics is the branch of robotics concerned with machines that work outdoors or in other unstructured, unprepared places: farms, mines, forests, oceans, the Arctic, disaster sites and other planets. A factory robot works in a cell built around it, with fixed lighting, known parts and fences. A field robot has to cope with the world as it is. This lesson sets out what makes that hard, how engineers think about environment, reliability and support, and how the rest of the course is organized. It assumes the localization and mapping methods of ROB-340, the planning methods of ROB-350 and the estimation methods of ROB-310, and builds on them rather than repeating them.

From the cell to the field

Industrial robots succeed because the environment is engineered to suit them. Parts arrive in fixtures, the floor is flat, nothing moves unexpectedly and the robot can repeat the same motion millions of times. Field robots invert almost every one of those assumptions. The ground is soft, wet, frozen or covered in rock; lighting changes from noon glare to darkness; rain, snow, fog and dust degrade sensors; vegetation grows and blows in the wind; animals, people and other machines appear without warning; and communication links come and go. The task itself is often less repetitive, because every field, ore face or ship hull is different.

Researchers often describe field environments along a few axes. Structure is how much of the environment was built for the robot or follows known rules. Dynamics is how much the environment changes while the robot works, from a slowly growing crop to a breaking wave. Harshness covers temperature, moisture, dust, pressure, radiation and vibration. Remoteness covers distance from human help and from reliable communication. A robot mowing a golf course is low on most axes; a robot inspecting a pipeline under Arctic sea ice is high on all four.

Lessons from field robotics history

Several landmark efforts shaped how the field approaches these problems. NASA's Sojourner rover in 1997, the Spirit and Opportunity rovers from 2004 and later Curiosity and Perseverance showed that careful engineering and conservative autonomy could keep machines working for years beyond their design lives on another planet; Opportunity, planned for 90 sols, operated for more than 14 years. The DARPA Grand Challenge of 2005, won by Stanford's Stanley, showed that a vehicle could drive 212 km of desert route without a driver by combining lidar, cameras and probabilistic methods. The DARPA Subterranean Challenge, which concluded in 2021, pushed teams of ground robots and drones to explore tunnels, caves and urban underground spaces with little or no communication; the systems competition was won by Team CSIRO Data61 from Australia. Commercially, autonomous haul trucks have operated in Australian iron ore mines since the late 2000s, and Imperial Oil reported in 2023 that its entire fleet of 81 haul trucks at the Kearl oil sands mine in Alberta was autonomous.

A common thread runs through these stories. Success rarely came from a single clever algorithm. It came from integrating robust sensing, conservative behavior, good fault detection, extensive field testing and an operations team that understood the robot's limits. Field robotics is as much systems engineering as it is algorithm design.

Environmental protection and hardening

Before any autonomy runs, the robot must survive its environment. Ingress protection is usually specified with the IP code of IEC 60529, in which the first digit rates protection against solids and dust and the second against water. An IP65 enclosure is dust-tight and protected against water jets; IP67 is dust-tight and survives temporary immersion; IP68 covers continuous immersion under conditions agreed with the manufacturer. Field robots usually need at least IP65 for outdoor work and IP67 or better where they may wade or be washed down. Connectors, cable glands, vents and seals around shafts are the usual failure points, not the enclosure walls.

Temperature is the second major stress. Electronics, lubricants, seals and especially batteries have limited operating ranges. Lithium-ion cells lose available capacity and power in the cold because their internal resistance rises, and most cannot be charged safely below about 0 degrees without heating. Many field robots therefore carry battery heaters, insulated enclosures or thermal management loops, and plan missions with derated energy budgets in winter. Heat is just as serious in deserts and mines, where solar loading and engine heat push electronics toward their limits. Vibration and shock from rough terrain loosen connectors, crack solder joints and fatigue brackets; designs use strain relief, locking connectors, potting and vibration testing to standards such as IEC 60068-2.

# Run time versus temperature for a field robot battery (illustrative derating curve)
import numpy as np
temps = np.array([-30, -20, -10, 0, 10, 20, 30])
usable = np.array([0.55, 0.70, 0.80, 0.88, 0.95, 1.00, 1.00])  # fraction of rated capacity
heater_w = np.where(temps < 0, 40 + 2 * (0 - temps), 0)       # heater power grows as it gets colder
capacity_wh, load_w, reserve = 2000, 250, 0.20
for t, u, h in zip(temps, usable, heater_w):
    hours = capacity_wh * u * (1 - reserve) / (load_w + h)
    print(f"{t:4d} C  usable {u:4.0%}  heater {h:3.0f} W  run time {hours:4.1f} h")

 -30 C  usable  55%  heater 100 W  run time  2.5 h
 -20 C  usable  70%  heater  80 W  run time  3.4 h
 -10 C  usable  80%  heater  60 W  run time  4.1 h
   0 C  usable  88%  heater   0 W  run time  5.6 h
  10 C  usable  95%  heater   0 W  run time  6.1 h
  20 C  usable 100%  heater   0 W  run time  6.4 h
  30 C  usable 100%  heater   0 W  run time  6.4 h

The table shows why winter missions need different plans from summer ones. The loss is not only stored energy: cold cells also deliver less peak power, which matters for robots that need bursts of current to climb slopes or to drive through deep snow.

Reliability, availability and redundancy

A field robot that fails far from help is expensive. Recovering a stuck rover in a muddy field costs labor and lost time; recovering a failed underwater vehicle may cost a ship-day worth tens of thousands of dollars; a failure on Mars cannot be repaired at all. Reliability engineering gives a vocabulary for these concerns. Mean time between failures (MTBF) is the average operating time between failures that stop the mission. Mean time to repair (MTTR) is the average time to restore the robot to service, including travel and diagnosis. Inherent availability is the fraction of time the robot is ready to work:

Field operations usually have a long MTTR because the robot is remote. That makes improvements in MTTR, such as modular swappable parts, remote diagnostics and spare robots, as valuable as improvements in MTBF. Where failures are random and the failure rate \( \lambda = 1/\mathrm{MTBF} \) is roughly constant, the probability that a robot completes a mission of length \( t \) without failure is \( R(t) = e^{-\lambda t} \).

Multi-robot systems change the reliability picture, which is one reason the second half of this course studies them. If a mission needs at least \( k \) of \( n \) robots to succeed, and each robot independently survives with probability \( p \), the mission succeeds with probability

# k-out-of-n mission reliability: how many robots must we send?
from math import comb
def k_of_n(k, n, p):
    return sum(comb(n, i) * p**i * (1 - p)**(n - i) for i in range(k, n + 1))
p = 0.90          # each robot completes the mission with probability 0.90
k = 4             # the survey needs at least 4 working robots
for n in range(4, 9):
    print(f"send {n} robots: mission success {k_of_n(k, n, p):.4f}")

send 4 robots: mission success 0.6561
send 5 robots: mission success 0.9185
send 6 robots: mission success 0.9841
send 7 robots: mission success 0.9973
send 8 robots: mission success 0.9996

Sending exactly the four robots needed gives only about a two-in-three chance of success, because any single failure ends the mission. Adding one spare raises success to over 90 percent, and two spares push it above 98 percent. The assumption of independent failures is optimistic: robots of the same design share software bugs, batteries from the same batch and the same weather, so common cause failures reduce the benefit, just as they do in the safety calculations of ROB-450.

Autonomy, supervision and the human team

Few field robots are fully autonomous in the sense of operating for long periods with no human involvement. Most work under supervised autonomy: the robot drives, plans and handles routine events itself, while a human supervisor monitors one or many robots, approves risky actions and intervenes when the robot is stuck or uncertain. ROB-430 lesson 6 described teleoperation and shared autonomy in detail; field operations add the problems of intermittent communication and of one supervisor responsible for a fleet. Lesson 11 returns to these issues. The human team also includes field technicians who maintain the robots, operations planners who decide where and when they work, and customers whose workflows change around the robots.

Field robotics programs often fail not because the robot cannot do the task in a demonstration, but because it cannot do it every day, in all weather, for a cost the customer can accept. A useful discipline is to measure mean distance or time between interventions, the fraction of the task completed without help, and the cost of each intervention. Those numbers, gathered over seasons rather than days, are what decide whether a field robot becomes a product.

How this course is organized

Lessons 2 and 3 cover the two foundations every ground robot needs outdoors: getting across rough terrain and knowing where it is when GNSS and lidar are degraded. Lessons 4 to 8 survey application domains: agriculture; mining, construction and energy; marine and underwater work; aerial field operations; and space. Each domain lesson highlights the physics and economics that shape its robots. Lessons 9 and 10 turn to teams of robots, first with explicit task allocation and then with swarms and distributed control. Lesson 11 covers communication, levels of autonomy and supervision in the field, and lesson 12 combines everything in the design of a multi-robot field mission. Regulation of mobile robots, automated vehicles and drones was covered in ROB-450 lessons 6 and 7, and automated vehicles in depth in ROB-420; this course refers to those lessons rather than repeating them.