Generational Robots

Persistent Autonomy

Adaptive epistemology requires learning across timescales that exceed any individual career, any institutional memory, any funding cycle. No human practitioner can span those timescales. This chapter proposes that generational robots, machines designed for habitat persistence rather than task completion, are the infrastructural answer to that temporal problem, not a fantasy but a trajectory already underway in environmental monitoring and conservation.

This chapter develops from a provocation I published eight years ago. “Designing Autonomy: Opportunities for New Wildness in the Anthropocene”, written with environmental historian Laura J. Martin and geographer Erle C. Ellis, and published in Trends in Ecology & Evolution in 2017, proposed that maintaining wild places in the Anthropocene paradoxically requires increasing levels of human management, and that this paradox might be resolved through fully automated systems capable of creating and sustaining wildness without ongoing direct human involvement. The paper introduced a conceptual design for a “wildness creator”, an autonomous landscape infrastructure system whose operations would eventually become “unrecognizable and incomprehensible to human beings,” producing ecological patterns divergent from anything humans had created (Cantrell, Martin, and Ellis 2017).

The paper was deliberately speculative. It surveyed eight existing semi-autonomous environmental management systems, from Oostvaardersplassen’s rewilding through surrogate megafauna to COTSbot’s autonomous removal of invasive crown-of-thorns starfish, and proposed that deep reinforcement learning systems, learning conservation strategies through environmental interaction rather than following pre-programmed rules, represented a viable extrapolation of trajectories already underway. But the paper left underspecified a set of questions that have become increasingly pressing as autonomous systems have matured. What is the temporal structure of a robotic system designed for generational operation? How does slow persistence change the epistemological character of what such systems can know? What does it mean to design landscapes that are maintained by systems whose decisions cannot be fully audited or understood?

This chapter takes those questions seriously. It draws on the robot ecology framework developed by Magnus Egerstedt, whose survivability constraints and habitat-centric design principles provide the engineering rigor that the Designing Autonomy paper gestured toward, to develop a more complete account of what generational robots would need to be, what they would need to know, and what landscape architecture’s role in designing their territories and ecologies might look like. The answer begins with an inversion, not speed but slowness.

Why Slowness Matters Now

The dominant cultural image of robots is still one of speed, expressed as agile quadrupeds sprinting across rubble, drones streaking through canyons, assembly arms flickering faster than the human eye. Speed has become synonymous with capability, and capability with value. But when we turn to environmental monitoring and landscape stewardship, speed is often a harmful temporal register. Forest dynamics, groundwater shifts, permafrost thaw, or coastal subsidence unfold over months, years, and generations. A robot that can sprint for twenty minutes is far less useful than one that can drift, amble, or hang motionless for years.

This temporal mismatch reveals something deeper about the relationship between speed, perception, and knowledge. Fast observation produces facts, discrete data points captured at high frequency, snapshots frozen in time. Slow observation produces something different, it produces possibility. Extended presence allows patterns to emerge, relationships to become legible, trajectories to unfold. The difference is not merely quantitative but epistemological. Fast sensing confirms or refutes hypotheses. Slow sensing generates them. Fast data closes questions. Slow data opens them.

Speed also determines the quality of interaction between observer and observed. A drone passing overhead at forty kilometers per hour captures an image but establishes no relationship with the landscape below. A robot dwelling in a marsh for months becomes entangled with that marsh and its movements synchronized to tidal rhythms, its energy budget coupled to solar angles, its sensors calibrated to the particular chemistry of that water. This entanglement is not contamination but the condition of co-habitation. The observer becomes part of the system observed, and that participation transforms what can be known.

This is the premise behind thinking about slow robots as a form of environmental infrastructure. Rather than one-off tools for short campaigns, these systems are conceived as enduring inhabitants of a landscape, moving just enough to track gradients, harvest energy, avoid hazards, and maintain vantage points, but otherwise tuned to the long rhythms of the systems they observe. In robotics, this shift shows up as a move from tasks accomplished over hours to long-duration autonomy as the core design objective. Magnus Egerstedt’s work on robot ecology formalizes this as a control problem in which robots must satisfy survivability constraints (maintaining energy, health, and safety) over open-ended time horizons, rather than simply reaching a stated goal (Egerstedt 2021). Slowness becomes a strategy and by moving less, consuming less, and accepting lower task throughput, robots can remain present long enough to matter for ecological processes.

Slowness also provides something that speed forecloses, a space for consideration. A fast agent reacts, a slow agent responds. The difference is consequential. Reaction is automatic, determined by pre-programmed thresholds and decision trees. Response involves interpretation, weighing, adjustment, a form of deliberation that unfolds in and requires time. When robots dwell rather than dash, they create temporal room for input from other agents, for evolution of conditions, for adaptation of strategies. They become participants in ecological time rather than interrupters of it.

At the landscape scale, this repositions robots from episodic visitors, like field crews, to something closer to resident species performing as agents with niches, diets, and temporalities that must be matched to the environments they inhabit. That ecological lens opens up a different set of design questions. How do robot populations share scarce recharging sites? How do they avoid over-exploiting their own habitats of energy and data? How do they coexist with, or even support, nonhuman organisms?

From Fast Agents to Robot Ecologies

Traditional robot design has followed a task-centric paradigm to design a robot capable of completing a job under expected conditions and evaluate success by task completion. In this mindset, the world is largely a backdrop and once the mission is done, the robot returns to a charging dock, a lab, or a scrap heap. The temporal horizon extends only to task completion, and success is measured in speed and efficiency.

Robot ecology inverts this framing. Instead of starting from a task, it starts from a habitat and asks what kinds of robotic agents could persist there over long durations, satisfying their survivability constraints while contributing a useful service (Egerstedt 2021). This ecology-inspired view emphasizes constraints over goals. Safety, energy, and health constraints must always hold, while task-related objectives such as coverage, data yield, or inspection frequency are pursued only within the space those constraints allow.

The shift from task to habitat transforms the temporal structure of robotic attention. Task-centric robots experience time as a countdown, a depleting resource to be optimized against mission objectives. Habitat-centric robots experience time as a medium, as an envelope within which presence, observation, and response unfold. This reframing changes what robots can perceive and know. The countdown (fast) perspective privileges facts with discrete measurements captured before the clock runs out. The medium (slow) perspective enables patterns to reveal relationships that only become visible through extended dwelling.

Robot ecology also shifts attention from individuals to populations. Long-duration autonomy is often easier to achieve with a population of robots that can distribute load, share resources, and compensate for failures than with a single heroic machine. Finally, it privileges niches over generic platforms, recognizing that robot ecologies benefit from specialized forms such as arboreal robots for canopy sensing, benthic crawlers for sediments, gliding autonomous underwater vehicles for the water column, and slow ground rovers for soil and vegetation.

The geomorphology modeling research conducted at the UVA (2017–present) provides a practice-based account of this transition from task-centric to habitat-centric design. The geomorphology table, a physical-digital hybrid environment of synthetic sediment, programmable water flow, and multi-modal sensing, was deployed initially through direct interaction, researchers physically placed objects, adjusted slopes, introduced sediment mixes, and observed outcomes. The system then moved toward responsivity, actuated sediment gates, sieves, and flow disruptors were programmed to react to sensed conditions, adjusting gate sequences based on detected sediment accumulation patterns. The research frontier at UVA moves toward the third mode, autonomy, how might the sediment gates develop response patterns through machine learning rather than pre-programmed rules?

The analogy is precise. A traditional thermostat operates through fixed threshold rules, when temperature drops below X, activate heating. A Nest thermostat develops patterns of automation through machine learning, optimizing across user schedule, energy pricing, and thermal dynamics, developing behaviors unique to the learning algorithm and the context in which it functions (Cantrell and Holzman 2016). The UVA research asks the equivalent question for sediment infrastructure, how might a gate system that learns from its own interactions with flow and deposition develop response patterns that a human designer could not have specified in advance? And what are the consequences, epistemological, ecological, ethical, when it does?

This gradient from interaction through responsivity to autonomy is not merely theoretical. It is a research trajectory the lab has been advancing for eight years, at experimental scale, before territorial deployment. The robot ecology framework that Egerstedt formalizes provides the vocabulary for understanding what each transition requires, the move from responsivity to autonomy demands not only different algorithms but different relationships between the robot and its habitat, longer time horizons, survivability constraints rather than task completion metrics, and the acceptance that the system’s behavior will evolve in ways its designers cannot fully predict or control.

Fast computation applied to slow ecological systems within a predictive frame produces a specific pathology. The algorithm responds at computational speed to processes that unfold across seasons and decades. Interventions adjusted daily on the basis of short-term fluctuations prevent slower processes from reaching the states that would have generated the most valuable knowledge. The system is managed into a narrow performance envelope that looks healthy on the dashboard while the biological community’s capacity to explore, to reorganize, to develop configurations no model anticipated, is progressively constrained. Generational robotics is designed against this pathology. Slowness is not a limitation. It is the temporal condition under which the territory’s own intelligence becomes legible.

This thinking is visible in the design of SlothBot, a wire-traversing robot developed to monitor environmental conditions in botanical gardens and forests. SlothBot hangs from cables, creeps slowly between trees, and relies on solar panels to harvest enough energy to remain operational for months. Slowness, in this case, is not a limitation but a carefully chosen adaptation to sparse energy and the need for low-disturbance presence in sensitive habitats (Egerstedt and Pauli 2020). The same logic underpins robotic concepts for persistent environmental monitoring and management in other domains. Long-range autonomous underwater vehicles (LRAUVs) designed for weeks-long scientific missions at sea prioritize energy efficiency, reliability, and the ability to dock, recharge, and re-task autonomously over dynamic performance (Ben-Yair et al. 2015).

Across these examples, robot ecology insists that we treat the robot-environment coupling as a mutual constraint satisfaction problem. The robot must satisfy its own survival constraints while remaining embedded in, and responsive to, the dynamics of the environment it inhabits.

Environmental Sensing as Robotic Habitat

Slow robots only make sense against the backdrop of rapidly evolving environmental sensing technologies. As sensor costs drop and power requirements shrink, the possibility of embedding dense networks of micro-sensors into landscapes, on posts, trees, buoys, building facades, or animal collars has expanded dramatically (Chan et al. 2020).

Two developments are particularly important for robot ecologies. First, low-cost, low-power sensing built on open-source platforms makes it possible to deploy large numbers of inexpensive sensors, provided calibration and drift are carefully managed. This democratizes data collection, enabling citizen science and distributed monitoring, but also produces new problems of data quality and standardization. Second, onboard computation and AI-enabled calibration allow machine learning models to correct biases and environmental interferences in low-cost sensor outputs, effectively bringing cheap hardware into research grade performance when coupled with reference data (Barcelo-Ordinas et al. 2022).

From a robot-ecology perspective, these trends mean that much of the environment is becoming a potential sensor substrate because robots are mobile and power-rich relative to static nodes and therefore can carry higher-quality instruments or serve as roving calibration platforms. In this emerging arrangement, robots and static sensors form complementary layers of an environmental nervous system. Static nodes provide continuous, localized time-series data, while slow robots weave among them, collecting more detailed measurements, cross-checking sensor readings, and filling spatial gaps. Edge computing on both nodes and robots allows “just enough” processing close to where data is produced, reducing bandwidth and enabling localized alerts.

The interaction between fast sensing and slow movement produces a particular mode of knowledge. Static sensors capture rapid fluctuations, the fast facts of momentary conditions. Mobile robots, moving slowly through sensor networks, integrate these facts across space and time, transforming discrete measurements into relational understanding. A temperature spike at one sensor becomes meaningful when the robot, having visited neighboring sensors, can situate it within a spatial gradient. A salinity reading becomes significant when the robot’s trajectory through time reveals whether conditions are trending toward or away from thresholds. The slow robot is not simply collecting data, it is also composing data into knowledge through the temporal structure of its attention.

A forest, for instance, might be equipped with a sparse grid of low-cost humidity and temperature sensors attached to tree trunks, while a wire-traversing robot moves slowly above the canopy, carrying more precise instruments, performing periodic calibration visits, and repositioning itself to chase anomalies. In short, environmental sensing turns landscapes into robot habitats where energy, data, and connectivity are spatially structured resources that robots must navigate and manage.

Algorithmic Cultivation (2019) encountered the generational robot problem in miniature, before the problem had a name. The robotic system was designed to operate for a year, the plants responding through growth configurations that the pruning would have incrementally shaped, the response becoming the input for subsequent pruning decisions. Technical difficulties limited the installation’s operation to weeks.

The failure is instructive in ways that success would not have been. Maintaining the coupled assemblage, keeping the robot’s sensing equipment calibrated in a humid growing environment, sustaining the data feeds, ensuring institutional support for an installation that would not resolve into a completed object, proved demanding in proportion to the biological agency the system was designed to cultivate. A simpler optimization-oriented system, programmed to maximize plant yield through precisely targeted irrigation and fertilization, would have been easier to maintain, its biological substrate was performing a specified role, not exploring one. The difficulty scaled with the degree to which the wetware was treated as a genuine participant.

Scaled to generational timeframes and territorial landscapes, these difficulties become the primary design problem. A slow robot dwelling in a marsh for years faces the Algorithmic Cultivation problem continuously, biofouling, sensor drift, component fatigue, and the biological dynamics of the environment it inhabits, which do not pause for maintenance cycles. The design priorities this chapter identifies, energetics as primary driver, reliability in the face of long-term exposure, quiet low-impact mobility, are not engineering preferences but responses to the specific failure modes that the Algorithmic Cultivation installation made visible at small scale. The installation’s truncated operation is the evidence base for why persistent presence is harder to achieve than it appears in speculation, and why the design of energetics, reliability, and maintenance protocols is as important as the design of sensing and actuation.

Slow Robots as Infrastructure

When slow robots are designed to remain in place for years, repeatedly traversing the same networks of wires, trails, currents, or air corridors, they begin to resemble infrastructure more than tools. Egerstedt describes this as treating robots as “persistent resources” embedded into the fabric of an environment rather than as disposable, mission-specific instruments (Egerstedt 2021). Thinking of robots as infrastructure highlights their persistence, dependability, and publicness.

Persistence means that the value of the system comes from continuous or regular presence, evident in always-on ocean gliders, canopy robots that never leave, barn guardians that monitor air quality around the clock. This persistence produces a form of temporal magnification that performs as an amplification of moments that would otherwise pass unregistered. A sensor that samples once per hour captures 8,760 measurements per year, a slow robot dwelling in the same location captures not just measurements but the texture of change between them, the rhythms and ruptures that constitute environmental time. Each moment is magnified because it exists within a continuous field of attention rather than as an isolated sample.

Dependability means that, like a bridge or a fiber-optic cable, a slow robot must be trusted to function in the background and that failure and maintenance cycles must be understood and planned for (Wang et al. 2019). Publicness, finally, acknowledges that many of the most consequential applications of climate monitoring, flood early warning, habitat surveillance have the character of public goods. Over time, persistent robots also develop a charismatic presence, becoming familiar features of the landscapes they inhabit and the species they share them with.

Examples of this infrastructural turn are appearing across domains. In agriculture, autonomous ground vehicles now crawl through fields carrying multi-parameter sensor suites, mapping moisture, temperature, and gas concentrations over large areas at fine spatial and temporal resolution. For tasks like seeding and weeding, robots reduce fuel use and operating costs compared to conventional machinery while enabling more precise environmental control (Tsanakas et al. 2023). In oceanography, fleets of LRAUVs and gliders, some able to dock autonomously at underwater stations, act as persistent sentinels of biogeochemical change, oil spills, and storm impacts, extending human observation into depths and durations that would otherwise be prohibitively expensive or dangerous (Ben-Yair et al. 2015). In forestry, platforms like SlothBot and newer wire-traversing robots such as RaccoonBot demonstrate how slow arboreal systems can instrumentalize tree canopies over long periods while relying on ambient energy and minimal human intervention (Pauli et al. 2020; Scherbat et al. 2025).

Across these cases, slowness is linked to energy autonomy and low disturbance. By minimizing motion and leveraging ambient energy (solar, wave, thermal, or even microbial fuel cells) slow robots can turn tenuous energy flows into operational lifetimes measured in months or years rather than hours (Pinto et al. 2019). For landscape architecture, conceiving of robots as infrastructure invites us to imagine corridors, nodes, and “robot-friendly” microarchitectures (gantries, mooring points, charging refuges) that are designed not just for human use, but for robot dwelling and circulation as well.

Designing for Slowness

Energetics, Reliability, and Quiet Mobility

If fast robots are engineered around actuation bandwidth and reaction speed, slow robots are engineered around energy budgets and failure modes. Three intertwined design priorities dominate, energetics as a primary driver, reliability in the face of long-term exposure, and quiet, low-impact mobility.

Energetics comes first. In robot ecology, energy is the ultimate limiting resource. Control policies are designed not only to achieve coverage or data collection objectives, but also to ensure that robots never violate survival constraints in that they must reach charging sites, stay within safe battery limits, and avoid behaviors that drain energy too quickly (Notomista and Egerstedt 2019). Slowness helps on multiple fronts. Lower actuation costs mean that moving slowly reduces dynamic losses and allows for lighter drive systems. Alignment with environmental processes means robots can dwell under sunlight, float in high-energy currents, or linger over microbial fuel cell patches, matching their duty cycles to ambient energy flows. Temporal flexibility, finally, reflects the fact that, because slow robots are not rushed to complete short missions, they can afford to pause in “feeding zones” to recharge. This is evident in systems like SlothBot, whose intermittent, deliberate movements are choreographed around solar charging cycles, and in LRAUV missions that plan sinuous, low-drag trajectories synchronized with ocean currents and docking opportunities (Pauli et al. 2020; Ben-Yair et al. 2015).

This energy-driven temporality creates a distinctive relationship to environmental time. Fast robots impose their schedules on landscapes where slow robots synchronize their schedules with landscapes. The sun rises, and the robot can move. Clouds gather, and the robot waits. This negotiation is not a limitation but a form of coupling and the robot’s rhythm becomes entrained to environmental rhythms, and this entrainment produces knowledge that speed would preclude. The robot learns, through its own energetic constraints, the temporal structure of its habitat.

Reliability emerges as the second priority. As robots stay in the field longer, failures cease to be isolated glitches and become systemic risks. Component fatigue, biofouling, sensor drift, and software bugs accumulate. Experience from long-lived service robots like TritonBot, deployed as a receptionist and tour guide, demonstrates that reliability engineering (health monitoring, fault diagnosis, and graceful degradation) is a major issue for long-duration autonomy (Wang et al. 2019). For environmental robots, the ecology metaphor suggests two complementary strategies. One is individual robustness where critical components are over-engineered, operating envelopes are kept conservative, and self-diagnostic and self-maintenance behaviors are incorporated, allowing robots to clean sensors, recalibrate, or switch to safe modes. The other is population redundancy where designers accept that individual robots will fail, but design the population to maintain function through overlapping coverage, role flexibility, and ease of replacement. Resilience is achieved not only through hardened hardware, but also through ecological design of robot communities.

Quiet mobility, the third priority, has ecological advantages. Slowness reduces disturbance to wildlife, decreases erosion, and lessens the chance of collisions or entanglement. Early field deployments of mobile robots in barns and poultry houses demonstrate that animals habituate to repeated, predictable movements of robotic platforms, integrating them into their behavioral routines (Berckmans 2017). For robots sharing space with sensitive species, deliberate, low-impact mobility will be as important as measurement capacity. In robot ecologies, behavioral compatibility with nonhuman inhabitants becomes a design criterion alongside energy and data.

Active Sensing and Robot Ecologies

Static sensors measure whatever happens to pass by, robots can choose where and what to measure. When combined with long-duration presence, this ability enables forms of active sensing that go beyond simple patrols.

In environmental monitoring, active sensing has emerged as a way to maximize information gain under constraints of time, energy, and communication. Gaussian process models, for example, allow robots to estimate spatial fields such as temperature, pollutant concentration, or algal density, and then select the next best measurement locations to reduce uncertainty (Chen et al. 2013). Recent work demonstrates how reinforcement learning and Bayesian optimization are being woven into active sensing strategies, especially for multi-robot teams exploring dynamic environments (Shi et al. 2024; Li et al. 2025).

Bringing this into robot ecology reframes active sensing as a kind of foraging behavior. Robots “graze” on information, moving toward regions where their marginal contribution to knowledge is highest. They must constantly balance exploitation, the intensive sampling of known hotspots, with exploration searching for new phenomena. Their movement patterns begin to resemble those of animals tracking patchy resources, subject to metabolic or energy constraints.

This foraging produces a distinctive epistemology. Fast surveys generate maps, comprehensive but static representations of spatial distributions at a moment in time. Slow foraging generates narratives of trajectories through space and time that reveal how conditions develop, how anomalies emerge, how systems transition between states. The slow robot’s knowledge is inherently temporal, and it cannot be extracted from any single snapshot but emerges from the accumulated experience of dwelling. This is knowledge as possibility rather than fact which is open to revision, sensitive to context, productive of questions rather than answers.

Over long periods, a population of slow robots may collectively learn seasonal routines, visiting wetlands during algal bloom seasons, patrolling particular ridges during fire season, or focusing on urban canyons during temperature inversions. These emergent patterns, encoded in protocols rather than static schedules, constitute a kind of robotic phenology that can be tuned and revised as conditions change.

For landscape architects and environmental managers, this opens up a different relationship to monitoring. Instead of specifying fixed transects or sample points, they can set high-level objectives and watch as the robot ecology organizes its own routines to meet them.

Multi-Robot Ecologies

Niches, Mutualisms, and Division of Labor

Robot ecology becomes most compelling when we think not of single devices but of robotic communities. Egerstedt’s framework highlights how different agents (ground rovers, aerial drones, underwater gliders, canopy climbers) can occupy distinct niches, each with its own diet (energy source), sensory capabilities, and mobility constraints (Egerstedt 2021). Within such communities, classic ecological relationships can be reinterpreted.

Mutualistic relationships appear when one type of robot provides a service that directly benefits another, and vice versa. For example, a solar-powered surface robot might act as a charging buoy for smaller underwater robots, in return, the underwater robots share subsurface data that improve the surface platform’s own decision-making about where to linger or move next (Song et al. 2021). Commensalism can describe arrangements where small low-power robots rely on the communication infrastructure provided by larger, more capable platforms, gaining connectivity without significantly affecting their hosts. Competition emerges when robots with overlapping sensing roles vie for access to scarce recharging stations, bandwidth, or other critical resources, requiring coordination protocols to avoid harmful interference.

These relationships are not just metaphors, they can be formalized as constraint-based control problems in which each agent’s behavior is governed by its own survivability constraints, task objectives, and coupling terms that encode cooperation or competition (Notomista and Egerstedt 2019). For environmental monitoring, multi-robot ecologies are particularly powerful when different modalities are needed. In a coastal wetland, a mix of fixed stations, slow rovers, and occasional aerial overflights can combine detailed process measurements with broader synoptic views. Rather than deploying separate, uncoordinated systems, a robot ecology approach treats these as interdependent species, co-designed to share energy, data, and infrastructure.

The temporal diversity of multi-robot ecologies amplifies their epistemic capacity. Fast-moving drones capture facts in the form of synoptic views, rapid assessments, and emergency responses. Slow-dwelling robots cultivate possibilities of emerging patterns, developing anomalies, gradual transitions. Together, they produce a layered understanding that neither could achieve alone where facts are situated within trajectories, snapshots contextualized by narratives.

When different robotic modalities produce conflicting accounts of the same patch of territory, that divergence is not a calibration problem to be resolved. It is the most valuable data the multi-robot system produces. A ground rover’s salinity reading that conflicts with an aerial drone’s NDVI reading of the same marsh is the territory communicating in two registers simultaneously. The chemistry of the water is saying one thing, the vegetation health is saying another, and the gap between them is where the most interesting ecological process is happening. Multi-robot ecologies are specifically designed to generate productive divergence of this kind, to hold multiple accounts of the same territory in tension rather than averaging them into consensus.

From Observation to Maintenance

The discussion so far has emphasized robots as observers sensing platforms that dwell, measure, and report. But the logic of slow robotics extends beyond observation to action. If robots can persist in landscapes long enough to understand their rhythms, they can also persist long enough to participate in their maintenance. This shift from sensing to stewardship represents a fundamental expansion of how robot ecologies can be imagined.

The precedents exist, though they are rarely framed in ecological terms. Agricultural robots that weed, thin, and prune are performing maintenance tasks calibrated to crop cycles. Autonomous underwater vehicles that remove invasive species or deploy restoration substrates are acting as ecological stewards or predators. What unites these examples is the coupling of persistent presence with deliberate intervention as robots that act because they have inhabited long enough to generate the knowledge to do so.

The temporal structure of slow maintenance differs fundamentally from fast interventions. A crew arriving to remove invasive vegetation works intensively for hours or days, then departs. The intervention is episodic, concentrated, and often disruptive an influx of heavy equipment, noise, compaction, the sudden absence of biomass. A slow robot performing the same function operates differently. It moves through the landscape continuously, removing individual plants as they appear, before they establish, before they seed. The intervention is distributed across time, woven into the ongoing life of the system rather than imposed as an external event. The cumulative effect may be equivalent, but the experiential and ecological texture is entirely different.

This distributed temporality reshapes how maintenance interacts with ecological processes. Slow robotic weeding allows native vegetation to respond gradually, filling gaps as they open rather than confronting sudden clearings. Slow robotic pruning shapes growth incrementally, guiding form through continuous small adjustments rather than periodic heavy pruning. Slow sediment management, robots repositioning material in small loads, day after day to mimic natural depositional processes rather than overriding them with dredges and dump trucks. In each case, the slowness of intervention allows the system to adapt, to incorporate the maintenance into its own dynamics, to remain coherent rather than being periodically disrupted and forced to recover.

The coupling of sensing and maintenance in the same robotic platform creates feedback loops unavailable to separated systems. A robot that both monitors vegetation health and removes invasive competitors can adjust its maintenance behavior based on what it observes, intensifying removal where natives are stressed, relaxing where they are thriving, pausing entirely during sensitive periods when birds are nesting. A robot that both measures sediment elevation and redistributes material can calibrate its interventions to observed accretion rates, adding sediment where the marsh is falling behind, holding back where natural processes are sufficient. This tight coupling of observation and action, mediated by algorithms but grounded in persistent presence, enables forms of adaptive management that would be impossible with separated sensing and maintenance systems operating on different schedules.

The implications extend to the character of the landscapes that emerge. Maintenance performed by slow robots produces landscapes that are continuously tended rather than periodically restored. The difference is visible and instead of the sawtooth pattern of degradation and intervention that characterizes conventional management, systems declining until a threshold triggers action, then being reset. Slow robotic maintenance produces smoother trajectories, systems held within desired ranges through ongoing adjustment. The variation occurs within a narrower envelope, guided by persistent attention rather than punctuated by a response to crisis.

There is also a question of what maintenance by slow robots does to the experience of landscape. Anthropogenic maintenance leaves traces of mown edges, pruned branches, repaired structures. These traces signify care, investment, the presence of tending hands. Robotic maintenance, especially when slow and distributed, leaves fewer visible traces. A marsh maintained by slow robots might appear untended, possibly even wild, spontaneous, self-organizing, even as it is continuously shaped by algorithmic intervention. This raises questions about legibility and honesty. Should robotic maintenance be visible, marked, acknowledged? Or is the appearance of autonomous wildness itself a legitimate design objective, provided the underlying care is genuine?

Multi-robot ecologies can distribute maintenance functions across specialized forms, just as they distribute sensing. One population might focus on vegetation management with slow crawlers that identify and remove invasive seedlings, trim encroaching growth, or disperse seeds of desired species. Another might focus on hydrological maintenance as surface craft that clear debris from culverts, adjust the position of floating structures, or redistribute sediment across tidal flats. A third might focus on infrastructure upkeep as robots that inspect, clean, and repair the cables, docks, and charging stations that support the robot ecology itself. This division of labor allows each form to be optimized for its particular task while the population as a whole maintains the landscape across multiple dimensions simultaneously.

The transition from observation to maintenance also changes the stakes of robot ecology. Observing robots that malfunction produce gaps in data where maintaining robots that malfunction may produce ecological harm. A weeding robot that misidentifies a native plant as invasive damages the system it was meant to protect. A sediment robot that deposits material in the wrong location may bury habitat or alter hydrology. The tolerance for error narrows as robots move from passive sensing to active intervention, and the requirements for reliability, interpretability, and human oversight correspondingly increase.

Robot Ecologies in Designed Landscapes

How might slow robots and robot ecologies become part of the practice of landscape architecture and territorial design?

Within an adaptive epistemology, landscapes are treated as learning systems, and design interventions are framed as hypotheses to be tested and revised (Cook 1999; Cantrell and Holzman 2016). Slow robots fit naturally into this framework as instrumental citizens, as an endemic species who watches, listens, measures, and reports back, supplying the feedback that allows designs to evolve. But they can also be active participants in the design’s realization, not just observing what happens but helping to make it happen.

A sediment-diversion landscape, for example, might host a small fleet of slow surface craft that continuously map bathymetry and suspended sediment, feeding data into morphodynamic models and triggering adjustments to gate operations. But those same craft, or others in the same ecology, might also redistribute sediment, nudging material toward areas where accretion is lagging, clearing channels that are silting in, reinforcing natural levees that protect interior marshes. The design specifies a set of targets and tolerances and the robot ecology works continuously to hold the landscape within those bounds as conditions evolve.

A permafrost park at the Arctic urban fringe might rely on slow ground robots to monitor ground temperature, microtopography, and vegetation change, adjusting boardwalk alignments or shading structures as thaw patterns shift. But those robots might also perform maintenance by repositioning thermal siphons, clearing drainage channels, removing vegetation that accelerates thaw, or installing reflective surfaces to reduce solar absorption. The line between monitoring and maintenance blurs and the robot becomes a continuous presence that both observes and tends.

Within the adaptive epistemology framework that this dissertation develops, robot ecologies are epistemological infrastructure, not only technological. Conventional monitoring infrastructure, fixed sensors, periodic surveys, remote sensing platforms, generates data about landscapes, the epistemological work of interpreting that data and revising design strategies is performed separately, by human practitioners operating on their own schedules and according to their own theoretical frameworks. Robot ecologies that both monitor and maintain collapse this separation. The robot’s sensing generates knowledge, the robot’s maintenance implements what that knowledge demands, the landscape’s response to that maintenance generates new knowledge that drives subsequent maintenance decisions. The feedback loop is closed within the robot ecology itself, operating at landscape timescales rather than institutional ones. Haque (2007), drawing on Pask’s cybernetic conversation theory, argues that responsive systems must be capable of mutual learning, not executing pre-programmed responses but evolving their behavior through sustained interaction with their environment. Robot ecologies that both sense and maintain achieve this at landscape scale.

This is the “learn-and-adjust” paradigm, central to the adaptive epistemology framework developed in Chapter 02, implemented as landscape infrastructure (Walters and Holling 1990; Holling 2004). The actuated sediment gates on the geomorphology table that develop response patterns through machine learning are prototyping, at experimental scale, what this looks like when the system learns faster than human practitioners can observe and specify. Programmed thresholds give way to learned behaviors as the system accumulates experimental data, not rules about what to do when turbidity exceeds X, but patterns about what interventions have produced, what outcomes in this table, with this sediment mix, under these hydrograph sequences. The machine learns the site through its performance. That is the shift from the thermostat to the Nest, from responsivity to autonomy, and at territorial scale, with decadal timescales, it is also the shift from a tool that executes a designer’s decisions to an infrastructure that generates its own. The generational robot does not implement the plan, instead it outlives the designer and continues learning after them.

The wildness creator described in “Designing Autonomy” (Cantrell, Martin, and Ellis 2017) is the limit case, a system whose learning has proceeded so far that its decisions are “unrecognizable and incomprehensible to human beings.” Between the Arduino-controlled gates of the Sedimachine and the wildness creator’s fully autonomous operations lies the design space that generational robots will inhabit, and that landscape architecture is uniquely positioned to shape.

The NEOM consultation is not the wildness creator. It is a conventional design consultation with a highly unusual client and an unusually ambitious brief, working within institutional constraints that require specified outcomes, deliverable milestones, and accountability structures that autonomous ecological management cannot satisfy. But it is the closest the practice has come to implementing the wildness creator’s underlying logic at territorial scale within an institutional framework that could actually build it. The adaptive canopy structures that self-fabricate and expand through self-made armatures, adapting to the life forming below, approach the wildness creator’s logic without achieving its autonomy. The internet of ecologies that relays real-time ecological data through local sensors approaches the sensing infrastructure the wildness creator would require. The consultation is the wildness creator theory in negotiation with institutional reality, finding the degree of autonomy that the institutional context can tolerate and building toward the fuller version that future institutional contexts might enable.

The Prototyping the Bay studio offers the pedagogical translation, framing design proposals as experiments, articulating what question the design is asking, and identifying what monitoring conditions would constitute evidence requiring design revision. This is adaptive epistemology as design practice. Robot ecologies are the infrastructure that makes this practice sustainable across the generational timeframes that landscape systems require, not replacing the designer’s judgment but extending the temporal reach of design attention into durations no human practitioner can personally inhabit.

The temporal magnification that slow robots provide transforms the designer’s relationship to what is designed. Traditional design operates through representation in drawings, models, and specifications that describe a future state and how it is to be constructed. Slow robots enable a different mode of operation that focuses on witnessing and tending. The landscape was already underway when the robot arrived. Its task is not to observe a passive system but to enter into relationship with a territory that has its own trajectories, its own timing, its own forms of communication. Witnessing is the practice of attending to what the territory is doing, not confirming what the design expected. They intervene to guide outcomes toward intentions, but the landscape decides what to do with those interventions. This witnessing produces knowledge that representation cannot, a knowledge of contingency, of surprise, of the gap between intention and outcome. And this tending produces landscapes that representation cannot, they are landscapes that are not built once but maintained continuously, held in dynamic equilibrium and at times ushered into new states through ongoing robotic care.

If robots are to inhabit designed landscapes for years and to maintain as well as monitor them, they need more than habitats, they need territories of operation. Overhead cable networks can support canopy robots and sheltered docking and charging alcoves can be integrated into piers, bridges, or berms or quiet service corridors can allow robots to move unseen and unobtrusively. But maintenance robots also need access to paths to reach every part of the site, clearances to perform their tasks, staging areas for materials they might deploy or remove. Designing such territories raises questions of coexistence and priority. Which spaces are for humans, which for robots, which shared? When do robotic maintenance operations pause to accommodate other uses? How are conflicts between robotic activity and human experience negotiated?

In some contexts, robots might be visible and celebrated as part of the landscape’s technological layer, explicitly expressing the monitoring and adaptive capacities of a site. In others, their operations might be tucked away to avoid disrupting human or nonhuman experiences of place. The risk of “robotic privilege,” design that caters primarily to robotic needs at the expense of other inhabitants, becomes more acute when robots are not just sensing but acting, reshaping the landscape in ways that may conflict with other values or uses. This is a spatial justice question as much as a technical one. A marsh maintained by slow robots whose docking infrastructure displaces the last accessible shoreline for a fishing community has not solved the maintenance problem, it has transferred it from the ecological to the social. The cable network that enables a canopy robot’s traversal may fragment the arboreal habitat it was deployed to monitor. The charging station that ensures a benthic crawler’s persistence may introduce electromagnetic noise that disrupts the organisms whose behavior the crawler is designed to observe. Landscape architects, with their expertise in spatial negotiation and multi-species occupancy, are particularly well positioned to choreograph these relationships, to ensure that the territories designed for robotic persistence do not become territories from which other inhabitants are excluded.

A landscape maintained by slow robots whose operations remain largely invisible raises a further question, one anticipated by Marenko’s (2022) concept of “hybrid animism,” where sensing surfaces and computational systems participate in landscape dynamics as quasi-agentive presences, that connects directly to the neo-wilds concept developed in Chapter 9, a continuously maintained landscape that appears wild, heterogeneous, dynamic, self-organizing, is a neo-wild. The computational management is present but operates at scales and temporal resolutions that render it imperceptible to human visitors. This is the paradox the dissertation traces, the wildest-appearing landscapes may be the most carefully tended. Slow robots are the mechanism through which that tending extends beyond human temporal capacity. They are the technological answer to the epistemological gap identified in Chapter 5, that computational knowledge cannot attend. Slow robots can.

Risks, Politics, and Ethics of Slow Robot Ecologies

The emergence of slow robotic infrastructure is not an unalloyed good. Many of the critiques leveled at smart-city systems and ubiquitous sensing apply here as well and the extension from observation to maintenance introduces additional concerns.

Persistent robots generate persistent data, and data governance quickly becomes central. Questions of ownership, access, and use are acute when robotic observations underpin decisions about zoning, insurance, compliance, or enforcement. Experiences with networked urban infrastructures demonstrate that optimization narratives often conceal power asymmetries, privileging the interests of platform providers and large institutions over local communities (Greenfield 2013). Robot ecologies will require governance frameworks that clarify who controls sensor payloads and data streams, provide mechanisms for communities to challenge, reinterpret, or supplement robotic observations, and ensure that environmental monitoring does not become a backdoor for surveillance of people or false narratives that allow for environmental exploitation.

When robots move from observation to maintenance, governance questions multiply. Who decides what maintenance is performed? What counts as an invasive species to be removed, a desirable trajectory to be supported, a threshold that triggers intervention? As with other concepts these are not merely technical questions but political ones, embedded in values about what landscapes should be and for whom. A robot ecology programmed to maximize habitat for migratory birds may conflict with one programmed to support recreational fishing and a maintenance regime optimized for carbon sequestration may suppress the biodiversity that local communities value. The algorithms that govern robotic maintenance encode assumptions about ecological priorities, and those assumptions deserve scrutiny and contestation.

The temporal structure of slow observation raises distinctive ethical questions. Fast data is ephemeral, captured and processed before its subjects are aware. Slow observation produces a different condition with an ongoing presence, accumulated records, evolving profiles of places and their inhabitants. This temporal magnification amplifies both benefits and risks. The benefits are epistemic, providing a richer understanding, more nuanced assessment, greater capacity for adaptive response. The risks are political creating deeper surveillance, more comprehensive control, less room for invisibility or escape. Slow maintenance adds another layer that is not just watching but shaping, not just recording but determining. The robot ecology becomes a governing presence, its algorithms inscribed in the landscape itself.

There are also ecological and social externalities. By design, slow robots maintain presence. Their presence or eventual absence is not neutral as they may alter animal behaviors, acting as de facto predators, competitors, or attractors and they may introduce new material flows (metals, plastics, battery chemistries) into sensitive environments and they may change labor dynamics, displacing field technicians, fishers, or rangers from roles long tied to local knowledge and livelihoods. When robots perform maintenance as well as monitoring, these displacements intensify. The roles of the gardener, the restoration ecologist, and the land manager are not eliminated but transformed, from direct tending to oversight of robotic tending. This shift changes the nature of environmental work, the skills it requires, and the relationships it sustains between people and place.

Egerstedt’s ecological framing helps us see these effects not as side issues but as part of the ecology itself where robots are new species entering already stressed systems. Ethical deployment demands careful study of these impacts and a willingness to redesign or withdraw robots where harms outweigh benefits (Egerstedt 2021). When robots actively maintain ecosystems, the stakes of getting this wrong increase. A misguided maintenance algorithm can cause lasting damage and a well-designed one can sustain valued conditions for decades. The difference lies in the quality of the values embedded in robotic behavior and the robustness of mechanisms for revising those values as understanding evolves.

A final concern is the autonomy-trust paradox. As robots become more autonomous, their decisions including where to go, what to measure, which alarms to trigger, and now, what maintenance to perform will shape environmental governance in ways that are difficult to audit. This paradox is already visible in other AI domains where the more capable and independent systems become, the harder it is for humans to understand or contest their actions. For slow robot ecologies that both observe and maintain, this raises practical and ethical questions. How do we verify that maintenance algorithms are not introducing biases or causing unintended harm? How do we ensure that robotic maintenance does not systematically neglect certain communities or habitats? What mechanisms exist for pausing or overriding autonomous behaviors when they conflict with other values?

Work on the Internet of Robotic Things emphasizes the need for secure, transparent, and manageable robot networks, but much remains to be done to translate those principles into environmental contexts (Palattella et al. 2024). Robust cybersecurity measures, explainable AI techniques, and independent auditing will be essential if slow robot ecologies are to become trusted components of environmental governance rather than opaque risk devices. When robots maintain as well as monitor, the requirement for transparency becomes more urgent and communities must be able to understand not just what robots observe but what they do, and why they do it.

Robotic Companions for Reflexive Stewardship

Slow robots and robot ecologies offer a way to weave machine intelligence into landscapes without defaulting to the logics of speed, disruption, or total control. When designed through an ecological lens that is attuned to energy flows, niches, coexistence, and long-duration survival they can become companions in the work of reflexive stewardship rather than overseers or intruders.

The epistemological contribution of slow robots lies in their slowness. They produce knowledge of a different kind than fast systems, not facts to be accumulated but possibilities to be explored, not snapshots to be catalogued but narratives to be interpreted. This knowledge requires time, the time for patterns to emerge, for relationships to develop, for the unexpected to reveal itself. Slow robots create that time by dwelling rather than passing through, by adapting rather than executing.

When slow robots extend from observation to maintenance, their contribution shifts from epistemological to ontological. They do not solely produce knowledge about landscapes, they participate in producing the landscapes themselves. This participation is continuous, distributed, and intimate and is woven into the daily life of the system rather than imposed upon it as an external force. The landscapes that emerge from such participation are different in kind from those produced by conventional design and maintenance as they are more responsive and more capable of evolving as conditions change.

The generational robot is not a tool imposed on a passive landscape. Blaise Agüera y Arcas argues that prediction and adaptation are not uniquely human or machinic and they are fundamental to all life, “everything alive is a computer” in the sense that life processes at every scale involve the processing of information, prediction of future states, and adaptation to feedback (Agüera y Arcas 2025). A slow robot navigating a wetland or monitoring a marsh edge is not imposing external control, it is participating in the computational processes the landscape already enacts. Generational robots are co-computational agents and neither masters nor slaves but participants in the distributed intelligence of the territory.

In the broader project of adaptive epistemologies, robot ecologies are best understood as infrastructures of attention and care. They extend our senses into places and times we cannot easily inhabit, providing the sustained, situated feedback required to treat design as an ongoing experiment. They extend our hands as well, performing the continuous small acts of maintenance that hold landscapes within specified trajectories. They challenge us to rethink what counts as infrastructure, who participates in monitoring and maintenance, and how we share responsibility for knowing and shaping our environments.

The temporal magnification that slow robots enable alters our relationship to the environments they observe and tend. Moments that would otherwise pass unregistered become legible, consequential, available for interpretation and response. Responses that would otherwise be impossible like the removal of a single invasive seedling before it establishes become feasible and woven into the ongoing metabolism of the landscape.

This is not merely data collection or task execution but a transformation of stewardship itself that creates a reorientation from the project with a beginning and end to the practice that persists across generations. Slow robots make this reorientation possible by providing the persistent presence, the sustained attention, and the gentle action that continuous stewardship requires.

The landscapes that emerge from such collaborations will likely be messier, more contingent, and more negotiated than static solutions. They will be places where wires and branches, buoys and marshes, algorithms and algae, humans and robots are all entangled in overlapping ecologies of care, constraint, and action. In that sense, slow robots are not just new tools, strangely they are new neighbors and new coworkers and learning to live and labor with them (critically, creatively, and justly) will be part of the design challenge ahead.

The generational robot acts with increasing autonomy by developing behaviors its designers did not specify and therefore operating outside human comprehension. But in what sense is this still design? Who is the author of a landscape that emerges from machine learning and biological agency symbiotically fused without direct human direction?