On a moonless night in late 2014, a small fixed-wing drone equipped with an infrared thermal imager lifted off from a ranger station in the Pretoriuskop section of Kruger National Park in northeastern South Africa, climbed to its operating altitude of roughly 100 meters, and began flying a programmed search pattern across roughly fifty square kilometers of scrub bush, dry riverbeds, and sparse miombo woodland. The drone’s pilot — a former park ranger named Graham Dyer, operating under a six-week trial contract — sat in front of a laptop in the ranger station, watching the thermal feed for the distinctive double signature that indicates a human figure on foot near a rhinoceros. The rhinoceros warms the savanna with the radiative pattern of a 3,000-pound mammal. The human shows up as a smaller, sharper, often-moving heat source against the same background, typically carrying a rifle. The drone records both signatures, transmits the coordinates back to the ranger station, and the on-foot patrol team is dispatched to interdict. Over the six weeks of the Pretoriuskop trial, while the drone was airborne, no rhinos were killed. In the previous month, in the same area, without the drone, nine rhinos had been poached.
This is the domain where the robotics industry’s environmental and conservation claims are stress-tested against the hardest possible measurement environment. Kruger National Park covers 19,485 square kilometers — roughly the size of Wales — and at the peak of the South African rhino-poaching crisis between 2013 and 2015, approximately 1,400 rhinos were being killed per year, an average of three to four per day. By 2020, that rate had fallen to one rhino killed approximately every 22 hours. By 2024, it had declined further, with a combination of armed patrols, dehorning programs, thermal-equipped drones, AI-based monitoring, and rhino relocation jointly responsible for the recovery. The conservation-drone fleet — Air Shepherd, a Lindbergh Foundation program that has flown over 4,000 missions across South Africa, Malawi, and Zimbabwe; the Hluhluwe/iMfolozi Park anti-poaching unit’s AI-and-thermal systems in KwaZulu-Natal; and a long tail of smaller park-specific deployments — is the most credible operational success story in the conservation-robotics category. The technology originally developed for U.S. military roadside-bomb detection in Iraq has been repurposed, with the same hardware family and the same image-processing algorithms, to do the exact opposite of what the autonomous-weapons industry is building it for — to detect humans who are about to kill animals, rather than to kill humans before they detect the drone.
The reforestation drone wave and the Mast pivot
In late 2016, a Seattle-based startup called DroneSeed — founded by Grant Canary, the CEO who had previously cycled through Techstars Seattle’s 2016 cohort — launched the most publicized application of robotics to climate-change mitigation that the industry had attempted: drone-swarm aerial reseeding of forested land destroyed by wildfire. The model was elegant on paper. The United States loses an average of 70,000 wildfires and 7 million acres of forest per year. Natural regeneration is slowing as wildfires get hotter and more frequent. Hand-planting reforestation crews are constrained by manual-labor scaling limits and a 2-to-3-year seedling supply chain bottleneck. A swarm of heavy-lift drones, each carrying a 57-pound payload of engineered “seed pucks” containing pine seeds, fertilizer, and a moisture-retention substrate, could in principle drop the supply chain bottleneck from 3 years to 3 months, plant tens of thousands of acres in days rather than seasons, and finance the whole operation through carbon credits sold to corporate buyers under the voluntary carbon market.
DroneSeed was the only reforestation company FAA-approved to fly drones with payloads above 55 pounds, to fly drones in swarms, and to fly drones beyond visual line of sight — a regulatory advantage that, in the parallel agricultural-drone market, would have been worth a significant valuation premium. The company rebranded as Mast Reforestation in 2023 (named for the forestry term mast years, the infrequent years when trees produce bumper crops of seed cones), acquired Silvaseed — a 130-year-old Western Washington seed bank that was the largest private seed supplier west of Colorado — in 2021, acquired Cal Forest Nurseries to become the largest seed-and-seedling vendor in the western United States, and built out a vertically-integrated pipeline that paired drone-deployed seed pucks with traditional hand-planted seedlings. By 2023, Mast had replanted approximately 2,500 acres of Montana and had a project pipeline of 20,000 additional acres. In February 2025, Mast closed a $25 million Series B round co-led by Chamath Palihapitiya‘s Social Capital, bringing total funding to roughly $81.74 million.
The operational results have, as of 2026, been substantially worse than the model predicted. In January 2025, Mast informed its partner Carbon Streaming that the drone- and hand-planted seedlings at the Sheep Creek, Baccala Ranch, and Feather River Dome projects had “experienced significantly higher than expected mortality rates and that the surviving seedlings had exhibited slower than expected growth rates.” Mast quietly withdrew several rounds of “forecasted mitigation units” — pre-sold carbon credits priced against the projected sequestration of planted seedlings — from the voluntary carbon market when the underlying biology failed to materialize. By June 2025, Mast was facing a fraud lawsuit from a former project partner. By February 2025, the company had pivoted its core business model from drone-and-hand reforestation to biomass burial — burying dead, fire-killed trees in clay-rich pits to prevent decomposition and trap their carbon underground — and announced the pivot alongside the Series B fundraise. The most ambitious conservation-robotics company of the 2016-2024 era is, in 2026, a tree-burial company that still does some drone seeding on the side. The promise the drones encoded — that you could mechanize reforestation at scale and finance it through carbon markets — has, structurally, not survived contact with the seedlings.
The post-Mast reforestation-drone ecosystem continues. Flash Forest in Canada operates a similar drone-seed-pod model focused on boreal reforestation. Dendra Systems (formerly BioCarbon Engineering), founded by ex-NASA engineer Lauren Fletcher, operates ecosystem-restoration drone projects in the United Arab Emirates, Australia, the United Kingdom, and Madagascar. AirSeed Technologies in Australia operates a drone-seed model focused on Australian native species and post-bushfire restoration. The combined deployed footprint is, by 2026, somewhere in the hundreds of thousands of acres treated cumulatively — a small fraction of the 70 million acres burned in the United States alone over the last decade, and a smaller fraction still of the global reforestation need. The technology works at the level of individual seed dispersal. The financial model that would scale it to the size of the problem has not yet emerged.
The anti-poaching drone and the night-vision arms race
The anti-poaching domain, by contrast, has been the conservation-robotics category with the cleanest operational evidence. Air Shepherd — formally part of the Charles A. and Anne Morrow Lindbergh Foundation — uses fixed-wing drones equipped with thermal-imaging cameras, originally developed for the U.S. military’s Iraq-era roadside-bomb-detection program, to fly nighttime patrols across high-poaching-probability zones in South African, Malawian, and Zimbabwean national parks. The drones operate primarily at night because approximately 80% of all poaching occurs in the hours of darkness. The thermal-imaging systems can distinguish the heat signature of a human carrying a rifle from the surrounding bush and animal heat. The on-the-ground response is conducted by armed park rangers; the drone is the detection layer, not the interdiction layer. As of 2026, Air Shepherd has operated over 4,000 patrol missions.
The operational impact, while difficult to attribute cleanly because the anti-poaching campaign has involved many parallel interventions (rhino dehorning, intelligence-led arrests, increased patrol funding, K-9 units, demand-reduction campaigns in Vietnam and China), is at minimum strongly correlated with a sustained decline in South African rhino mortality. The peak of approximately 1,400 rhinos poached per year in 2014 had declined by roughly 60-70% by 2024. Crawford Allan, the World Wildlife Fund’s crime-technology project spokesperson, has publicly described Kruger as “ground zero for poachers,” with as many as twelve organized poaching gangs operating inside the park at any given time. The conservation-drone fleet has, in the operational reading of the WWF and the South African National Parks (SANParks) leadership, contributed materially to the reduction. The same family of camera-and-autonomy technology that runs the DFR drone programs at Chula Vista PD is, in Kruger, watching rhinoceroses sleep — a structural reuse of the same Skydio and DJI-derived platform stack that has scaled into every other drone-deployment domain in the cluster. The hardware stack depends on the same semiconductor supply chain, the same lithium-ion battery chemistry, and the same rare-earth permanent magnets in the motors as every other autonomous platform the cluster has documented — including the same Boston Dynamics Spot platforms that several South African private game reserves have, since 2024, begun acquiring for perimeter patrol and night-time inspection of remote ranger outposts.
The conservation-drone story extends well beyond anti-poaching. South African conservationist Carel Verhoef in 2024 used a small fleet of drones and ranger pilots to move a herd of 150 elephants 70 kilometers at night across the Tanzania-Kenya border, using the drones as a noise-and-presence shepherding tool to redirect the herd away from a corridor where they were vulnerable to poaching and toward a protected reserve. Chisl/Veriphy AI, a Johannesburg-based group founded by Willem Kellermann, conducted a major drone-based wildlife census in 2025 covering more than 100,000 hectares in several private game reserves near Kruger, using AI-driven image processing to count elephants, rhinos, buffalo, antelope, and lions at a fraction of the cost of historical helicopter-based aerial census methods. Ezemvelo KZN Wildlife in KwaZulu-Natal flies BVLOS drones for both rhino-monitoring and rare-plant work — including a multi-year project to locate the so-called “loneliest plant in the world,” a single specimen of Encephalartos woodii believed to be the last of its species, using a combination of drones, satellites, and spectral imaging. The conservation-drone footprint across sub-Saharan Africa is, by 2026, somewhere in the low thousands of operational airframes across hundreds of parks and reserves.
RangerBot and the Great Barrier Reef
In August 2018, after winning the $750,000 People’s Choice prize at the 2016 Google Impact Challenge, researchers from Queensland University of Technology under principal investigator Matthew Dunbabin launched RangerBot at the Reef HQ Aquarium in Townsville, Queensland. RangerBot is a 15-kilogram autonomous underwater vehicle with six thrusters, two stereo camera systems for visual navigation, and a single dedicated function: identify and inject the crown-of-thorns starfish (COTS), the invasive coral-eating echinoderm whose population booms across the Great Barrier Reef have, since the early 2010s, been one of the most consequential drivers of coral loss after thermal bleaching. RangerBot identifies COTS with 99.4% accuracy using onboard computer vision, dispatches a lethal dose of vinegar or bile salts via injection arm into each identified specimen, and operates for eight hours on a single charge — roughly three times longer than a human diver can stay below the surface.
The structural argument for RangerBot was scale economics. The Great Barrier Reef Marine Park Authority (GBRMPA) reported that across 2023-2024, 16,657 hours of human-diver effort culled approximately 50,227 COTS — a rate of one starfish killed every 20 minutes. A fleet of RangerBots, each operating continuously for eight-hour shifts and identifying COTS in real time, could in principle achieve culling rates an order of magnitude higher than the diver-based baseline. The actual operational deployment, as of 2026, remains in the low-single-digit-fleet-size range — RangerBot is built in QUT laboratories rather than mass-produced by a commercial manufacturer, and the GBRMPA’s COTS-control program remains predominantly diver-based. The complementary Down Deep Drones prototype, built by an independent Australian developer for approximately $6,000 on an off-the-shelf QYSEA underwater drone platform, has been pitched to GBRMPA and the Reef and Rainforest Foundation since 2018 with mixed reception. The technology works on a per-starfish basis. The institutional adoption pathway that would scale it to the size of the COTS outbreak has not closed.
The broader underwater-conservation-robotics ecosystem includes LarvalBot (a sister project at QUT that dispenses coral larvae onto degraded reefs to accelerate regeneration), Mesobot at the Monterey Bay Aquarium Research Institute (which tracks individual zooplankton at midwater depths for ocean-research purposes), and a growing fleet of academic-research AUVs operating in the same family of low-cost commercial platforms — OpenROV Trident units, QYSEA FIFISH professional models, and the Blue Robotics BlueROV2 — that have made underwater robotics accessible to research budgets that could not previously afford an oceanographic-grade ROV. The combined deployed footprint of conservation-and-research AUVs across global coastal-management programs is, by 2026, in the low tens of thousands of units, dominated by the consumer-grade Chinese platforms and the academic-grade U.S. and European systems.
Forest inventory, LiDAR drones, and the timber supply chain
The commercially largest application of drones in the broader land-management category is forest inventory — the cataloging of standing timber, biomass density, species mix, and harvestable volume across managed and unmanaged forests for the timber, paper, carbon-credit verification, and forest-management industries. Treeswift, a Philadelphia-based startup, operates a fleet of LiDAR-equipped autonomous drones that fly under forest canopy to inventory individual trees, identify species, and measure trunk diameter at breast height — work that historically required ground crews with handheld measuring tape and clipboards. Sweden’s Skogforsk research institute operates a comparable program for the Scandinavian timber industry. Finland’s Metsähallitus flies drones for state-forest inventory. The U.S. Forest Service operates several thousand drones across its 193-million-acre management portfolio for fire-line monitoring, post-fire assessment, invasive-species surveys, and recreation-area management. The Bureau of Land Management operates a parallel fleet across the 245 million acres under its jurisdiction.
The forest-fire-monitoring side of the land-management domain bleeds directly into the autonomous wildfire-suppression aircraft documented in the firefighting cluster post — the Sikorsky-Rain autonomous Black Hawk that conducted live-fire suppression tests in April 2025 is, in operational terms, the upper end of the same fire-monitoring-and-suppression continuum that smaller drone fleets occupy at the lower end. Pano AI, a San Francisco startup that operates a network of high-mountain cameras for early wildfire detection, integrates with state-fire-agency drone-dispatch systems across California, Oregon, Washington, Colorado, and several other Western states. The combined real-time wildfire-monitoring fleet across the U.S. West — drones, fixed cameras, satellite-based hot-spot detection, and crewed reconnaissance aircraft — has dramatically reduced the average time between fire ignition and first response over the last decade, with corresponding measurable reductions in average burn area for fires detected in the first hour.
The wildlife census and the disappearing penguin
The most consistently funded and operationally successful category of conservation drone is the wildlife population census. The British Antarctic Survey has, since 2017, used fixed-wing drones to count penguin colonies across the Antarctic Peninsula, South Georgia, South Orkney, and the South Sandwich Islands — work that historically required ship-based expeditions counting from binoculars and which the drones now accomplish in fractions of the time at a fraction of the cost. The University of Sydney‘s wildlife-monitoring drone program counts kangaroo, wallaby, and koala populations across New South Wales and Queensland. The U.S. National Park Service flies drones for Yellowstone bison counts, Glacier bighorn sheep counts, and Channel Islands fox monitoring. The University of Cape Town‘s African Penguin Initiative uses drones to count breeding colonies along the South African coast — a population that has, despite the monitoring, declined by more than 60% since 2000 and is now classified as critically endangered.
The structural distinction in the wildlife-census category is that the drone is a measurement instrument rather than an intervention. The robot does not change the population. It tells the conservation manager what the population is. The decisions about whether to relocate animals, install electric fencing, deploy anti-poaching patrols, or close fisheries to protect prey species are downstream of the data. The conservation outcome depends on the institutional capacity to act on the measurement. This is the recurring constraint in every conservation-robotics deployment the cluster has documented — the robots can do the surveillance and the intervention, but the conservation outcome depends on the political, legal, and financial framework around them. The Air Shepherd drone identifies the poacher. The on-foot ranger team has to make the arrest. The RangerBot identifies the COTS. The GBRMPA management plan has to scale the deployment. The Mast Reforestation drone drops the seed puck. The seedling has to survive the next three drought summers.
Marine conservation and the Saildrone fisheries program
The largest operational deployment of autonomous vehicles in marine conservation in 2026 is the NOAA Fisheries program that uses Saildrone Voyager units for trawl-survey calibration, salmon-population assessments off the U.S. West Coast, pollock-population assessments in the Bering Sea, and acoustic monitoring of cetacean populations across the U.S. EEZ. Saildrone has, as of 2026, completed multi-year contracts with NOAA, with the U.S. Coast Guard for civilian and dual-use missions, and with the Australian Bureau of Meteorology for Pacific climate monitoring. The vessels are the same 23-foot solar-and-wind-powered platforms that the maritime defense industry has scaled for U.S. Navy task force operations — the dual-use overlap is total. The same Voyager that maps a Bering Sea pollock biomass survey in March can be re-tasked for Replicator maritime-domain-awareness in the South China Sea in June with no hardware modification.
The fisheries-assessment use case is, in conservation-robotics terms, the strongest published-evidence example outside of African anti-poaching. NOAA’s Saildrone-based pollock surveys have, in head-to-head comparison studies against traditional crewed fishing-vessel-and-acoustic-transducer assessments, produced comparable biomass estimates at substantially lower cost and with substantially less impact on the surveyed fish populations. The structural argument for the autonomous platform is the same as it is in every other robotic-deployment domain in the cluster: the unit cost is lower, the duration is longer, the human risk is lower, and the data quality is, in some categories, measurably better.
What 2026 looks like in conservation robotics
In 2026, Air Shepherd’s anti-poaching drones continue to fly across South Africa, Malawi, and Zimbabwe, with the broader anti-poaching technology ecosystem — thermal imaging, AI-driven image processing, BVLOS regulatory waivers, integrated ranger dispatch — credited with material contribution to the ~60-70% decline in South African rhino mortality since the 2014 peak. Mast Reforestation continues to operate as a hybrid drone-seeding-and-biomass-burial business, with the original drone-swarm reforestation model having largely failed against its carbon-credit projections, and a fraud lawsuit pending against the company. Flash Forest, Dendra Systems, and AirSeed Technologies continue to operate smaller reforestation-drone programs in Canada, the UAE/Australia/UK/Madagascar, and Australia, respectively. RangerBot continues to be deployed in limited fleet sizes at the Great Barrier Reef alongside the larger diver-based COTS-control program. Treeswift, the U.S. Forest Service, the Bureau of Land Management, and a constellation of state and private timber-industry operators run a forest-inventory drone fleet measured in the low tens of thousands of airframes. NOAA’s Saildrone fisheries-assessment program continues to expand. The British Antarctic Survey, the U.S. National Park Service, and a long tail of academic wildlife-census programs continue to operate drone-based population counts that have replaced helicopter-and-binocular-based methods at orders-of-magnitude lower cost.
The conservation-robotics category does something the rest of the cluster has not asked the technology to do — it asks the robot to be the substitute for institutional capacity that the conservation movement has not been able to build at scale. The reforestation drone was supposed to replace the manual planting crew that the forestry industry cannot afford to scale. The anti-poaching drone was supposed to replace the ranger patrol that the African national parks cannot fund to the size of their territories. The RangerBot was supposed to replace the human diver who cannot stay submerged long enough to keep up with the COTS outbreak. The wildlife-census drone was supposed to replace the helicopter survey that no national park system in the world has budgeted at the frequency the science requires. In each case, the robot does the work the human alternative cannot do — and in each case, the binding constraint on the conservation outcome is not the robot’s capability but the institutional structure around it. The carbon-credit market has not been able to verify the Mast Reforestation projects’ biological outcomes. The South African rhino population is recovering not because the drone alone interdicts the poacher, but because the drone’s detection feeds an armed ranger team that the South African government has been willing to staff and arm at scale. The Great Barrier Reef’s COTS population is not falling fast enough because the RangerBot fleet is not big enough, because the GBRMPA budget is not large enough, because Australian climate policy has not, in the operational reading of the marine-biology community, addressed the underlying nutrient-runoff and thermal-bleaching pressures that drive the COTS outbreak in the first place.
The robots in this cluster are, in some ways, the cluster’s most morally compelling deployments — the Spot patrolling an offshore oil platform is not saving an endangered species, the Trajekt Arc throwing 100-mph cutters in a basement batting cage is not buying time for a coral reef, and the humanoid robot demoing on a stage at a venture-capital conference is not, in any direct sense, addressing the biosphere collapse that the conservation-robotics community has spent the last fifteen years building hardware against. The conservation drone, the anti-poaching thermal imager, the reforestation seed puck, the underwater starfish-injector, and the autonomous fisheries-assessment platform are the rare robots whose mission statement is, structurally, “do something the planet’s biosphere desperately needs.” The fact that the conservation-robotics category has the most ambitious mission and the most mixed operational evidence is not, in the cluster’s running thesis, a failure of the robots. It is a failure of the institutional framework around the robots — the carbon markets, the national park budgets, the international wildlife-trade enforcement regimes, the climate-policy frameworks, the conservation-infrastructure budgets that no national government has been willing to fund at scale — to match the capability of the underlying robotic platforms the scientific research community and the K-12-to-university talent pipeline have spent decades producing — including the deliberately-cute consumer-facing robots whose design budgets, in 2026, dwarf the entire global conservation-robotics R&D spend by a factor of perhaps fifty to one. The robots will keep doing the work. Whether the planet’s ecosystems recover enough to justify having built them is, in 2026, still being decided by the institutions the robots cannot, by themselves, fix — and the gap between the robotic capability and the conservation outcome remains, across every domain the cluster has documented, the most morally consequential and the least technologically solvable problem in the entire industry.