The Indian workers training AI robots to take their jobs
AFP
June 11, 2026
AFP
June 11, 2026
Indian housewife Nagireddy Sriramyachandra wears a smartphone on her head as she records her actions through motion capture while slicing mangoes at her home in Chennai – Copyright AFP R.Satish BABU
With a smartphone strapped to her head, Indian housewife Nagireddy Sriramyachandra films herself slicing mangoes to train AI-powered robots to take on household jobs in the future.
Earning just over two dollars for an hour of video, her mundane recordings are invaluable for global tech companies teaching machines how to move like humans in the real world.
The 25-year-old is one of a growing army of thousands of AI system trainers in the world’s most populous country.
“Who else will give you 250 rupees an hour just for doing housework?” said Sriramyachandra from her kitchen in Chennai in southern India’s Tamil Nadu state.
“I may get a robot myself in the future,” she added.
Artificial intelligence chatbots and image generators crunch reams of digital data, but building systems to navigate real-life environments is more challenging.
Developers think feeding first-person footage, called “egocentric data”, into specialised AI models will help robots copy humans.
Some AI trainers work at home, others in factories or specialised studios — using video glasses, head-mounted cameras and motion sensors.
“It blares ‘hands not detected’ when I’m not recording properly,” said Sriramyachandra, who sends recordings via a special app to the AI data company Objectways.
The firm, which has offices in India and the United States, lists Fortune 500 multinationals as clients. It works with Amazon SageMaker, a platform for machine learning models.
– ‘Better things’ –
The humanoid robot market is booming, with investment bank Morgan Stanley predicting there could be over a billion in use by 2050, mostly for industrial and commercial purposes.
“Folding clothes, coffee making… cooking a very specific thing, sandwich making,” Objectways head Ravi Shankar said, listing videos requested by clients.
“Some jobs are supposed to be taken over, so humans can go and do better things.”
In India, the emerging field of spatial AI is providing new employment — for now.
The 50-year-old CEO is US-based, but hires workers from Tamil Nadu, where he grew up, one of India’s international technology hubs.
At a Karur textile factory, busy with workers attaching labels to caps and ironing cloth bags, AFP saw eight people wearing head cameras and smart glasses supplied by Objectways.
India has positioned itself as a global middleman for the creation, processing and annotation of AI data.
“It’s likely that these data collection services will increase”, said digital labour expert Aditi Surie, from the Indian Institute for Human Settlements in Bengaluru.
– Informal workers –
India is aggressively developing its AI industry, but its leaders are aware that, alongside the technology’s much-hyped benefits, automation poses risks.
Government think-tank NITI Aayog said that most discussions around artificial intelligence and labour “focus on white-collar professionals and predict an almost certain loss of jobs in the segment” without urgent action.
“Little attention, if any, is paid to how AI can serve India’s 490 million informal workers, the very people who form the backbone of our economy,” it said in a report released ahead of a global AI summit in India this year.
The think-tank has examined how the technology could help or harm dozens of professions — from cobblers to sewer cleaners, farmers to tea sellers.
For the last decade, 55-year-old Ponni has sat on a roadside in Bengaluru, the city known as India’s Silicon Valley, making flower garlands.
She, too, has been paid to have a phone strapped to her forehead.
“The next generation… who might have to do work similar to mine — they will face a problem,” Ponni said.
– Always wearing a camera –
At an Objectways studio, AI system trainers film themselves performing household tasks in fake, fully furnished apartment rooms.
After several thousand hours of filming, the wallpaper is changed to provide clients with variety.
“Today I sit here, tomorrow I stand there,” said engineering graduate Rani N., 21, on a break from filming herself, once again, folding a towel.
Each video lasts about four minutes, and she records around 90 a day — on nearly every conceivable spot on the bed.
She says the job is “tolerable”, but feels like she’s always wearing a camera.
In other rooms, colleagues arranged pencil sharpeners, water bottles and crayons in patterns, recording with depth-sensor cameras.
Qanat Consulting Services in Andhra Pradesh, an Objectways subcontractor, supplies about a dozen larger data firms with recordings.
Some of its 2,000 contributors perform tasks with motion-sensor bands on their “wrists, hands and legs”, CEO Thaslim Pattan said.
Manish Agarwal of Bengaluru-based Humyn Labs, not related to Objectways, records conversations as well as videos.
Contributors discuss assigned topics — ranging from politics to entertainment — for clients wanting to process speech patterns.
Agarwal denies that robots will steal jobs, believing that networks of humans and robots “will work together” one day, he said.
“A welder in India could be managing a welder-robot in Prague,” he said.
With a smartphone strapped to her head, Indian housewife Nagireddy Sriramyachandra films herself slicing mangoes to train AI-powered robots to take on household jobs in the future.
Earning just over two dollars for an hour of video, her mundane recordings are invaluable for global tech companies teaching machines how to move like humans in the real world.
The 25-year-old is one of a growing army of thousands of AI system trainers in the world’s most populous country.
“Who else will give you 250 rupees an hour just for doing housework?” said Sriramyachandra from her kitchen in Chennai in southern India’s Tamil Nadu state.
“I may get a robot myself in the future,” she added.
Artificial intelligence chatbots and image generators crunch reams of digital data, but building systems to navigate real-life environments is more challenging.
Developers think feeding first-person footage, called “egocentric data”, into specialised AI models will help robots copy humans.
Some AI trainers work at home, others in factories or specialised studios — using video glasses, head-mounted cameras and motion sensors.
“It blares ‘hands not detected’ when I’m not recording properly,” said Sriramyachandra, who sends recordings via a special app to the AI data company Objectways.
The firm, which has offices in India and the United States, lists Fortune 500 multinationals as clients. It works with Amazon SageMaker, a platform for machine learning models.
– ‘Better things’ –
The humanoid robot market is booming, with investment bank Morgan Stanley predicting there could be over a billion in use by 2050, mostly for industrial and commercial purposes.
“Folding clothes, coffee making… cooking a very specific thing, sandwich making,” Objectways head Ravi Shankar said, listing videos requested by clients.
“Some jobs are supposed to be taken over, so humans can go and do better things.”
In India, the emerging field of spatial AI is providing new employment — for now.
The 50-year-old CEO is US-based, but hires workers from Tamil Nadu, where he grew up, one of India’s international technology hubs.
At a Karur textile factory, busy with workers attaching labels to caps and ironing cloth bags, AFP saw eight people wearing head cameras and smart glasses supplied by Objectways.
India has positioned itself as a global middleman for the creation, processing and annotation of AI data.
“It’s likely that these data collection services will increase”, said digital labour expert Aditi Surie, from the Indian Institute for Human Settlements in Bengaluru.
– Informal workers –
India is aggressively developing its AI industry, but its leaders are aware that, alongside the technology’s much-hyped benefits, automation poses risks.
Government think-tank NITI Aayog said that most discussions around artificial intelligence and labour “focus on white-collar professionals and predict an almost certain loss of jobs in the segment” without urgent action.
“Little attention, if any, is paid to how AI can serve India’s 490 million informal workers, the very people who form the backbone of our economy,” it said in a report released ahead of a global AI summit in India this year.
The think-tank has examined how the technology could help or harm dozens of professions — from cobblers to sewer cleaners, farmers to tea sellers.
For the last decade, 55-year-old Ponni has sat on a roadside in Bengaluru, the city known as India’s Silicon Valley, making flower garlands.
She, too, has been paid to have a phone strapped to her forehead.
“The next generation… who might have to do work similar to mine — they will face a problem,” Ponni said.
– Always wearing a camera –
At an Objectways studio, AI system trainers film themselves performing household tasks in fake, fully furnished apartment rooms.
After several thousand hours of filming, the wallpaper is changed to provide clients with variety.
“Today I sit here, tomorrow I stand there,” said engineering graduate Rani N., 21, on a break from filming herself, once again, folding a towel.
Each video lasts about four minutes, and she records around 90 a day — on nearly every conceivable spot on the bed.
She says the job is “tolerable”, but feels like she’s always wearing a camera.
In other rooms, colleagues arranged pencil sharpeners, water bottles and crayons in patterns, recording with depth-sensor cameras.
Qanat Consulting Services in Andhra Pradesh, an Objectways subcontractor, supplies about a dozen larger data firms with recordings.
Some of its 2,000 contributors perform tasks with motion-sensor bands on their “wrists, hands and legs”, CEO Thaslim Pattan said.
Manish Agarwal of Bengaluru-based Humyn Labs, not related to Objectways, records conversations as well as videos.
Contributors discuss assigned topics — ranging from politics to entertainment — for clients wanting to process speech patterns.
Agarwal denies that robots will steal jobs, believing that networks of humans and robots “will work together” one day, he said.
“A welder in India could be managing a welder-robot in Prague,” he said.
Struggling German auto supplier Bosch pivots to robots
AFP
June 10, 2026
AFP
June 10, 2026
Bosch says the market for its sensors that are used in robots is growing rapidly – Copyright AFP/File RONALDO SCHEMIDT
German industrial giant Bosch said Wednesday it will step up efforts in the field of humanoid robotics, as its traditional auto parts business comes under increasing pressure.
The world’s biggest auto supplier, Bosch makes everything from braking systems to sensors, but has suffered as European carmakers battle fierce overseas competition and weak demand.
However, the rise of humanoid robots, powered with generative AI models and capable of performing complex tasks, offers an opening for the group, chief executive Stefan Hartung said.
“With the advent of humanoid robotics, the demand for Bosch components and solutions is increasing,” he said in a statement.
The market for specialised MEMS sensors is expected to grow to over $19.2 billion by 2030 and hit an annual growth rate of four percent, according a study by consultancy Yole Group, which was presented by Bosch.
Bosch is a key producer of the tiny sensors, which are crucial in robotics.
At an event in Berlin, Hartung stressed the importance of the components in improving the dexterity of robots.
These sensors determine whether the robot “should tighten its grip or not, whether it is dealing with a sturdy object, or whether it needs to act delicately because it is an egg,” he said.
“Humans have four million touch sensors. If we were to build robots equipped with as many sensors, four years of global sensor production would barely be enough to equip 12,500 robots,” he added.
The focus on automation is also meant to boost the competitiveness of Bosch’s German factories and plug shortages of skilled labour.
Bosch, also known for making a wide range of industrial equipment and household appliances, struck a deal with German robotics firm Neura in January to gather data on factory work.
Under the partnership, several thousand workers in some of Bosch’s 350 facilities worldwide will wear sensor suits to glean training data for Neura robots.
Neura said Wednesday it had raised up to $1.4 billion in fresh fundraising from backers including Bosch, chip giant Nvidia, Amazon and crypto group Tether.
The funding will be used to accelerate Neura’s activities, ranging from the deployment of robots to the rollout of “gyms” for clients to train bots, and the development of the company’s physical AI systems.
German industrial giant Bosch said Wednesday it will step up efforts in the field of humanoid robotics, as its traditional auto parts business comes under increasing pressure.
The world’s biggest auto supplier, Bosch makes everything from braking systems to sensors, but has suffered as European carmakers battle fierce overseas competition and weak demand.
However, the rise of humanoid robots, powered with generative AI models and capable of performing complex tasks, offers an opening for the group, chief executive Stefan Hartung said.
“With the advent of humanoid robotics, the demand for Bosch components and solutions is increasing,” he said in a statement.
The market for specialised MEMS sensors is expected to grow to over $19.2 billion by 2030 and hit an annual growth rate of four percent, according a study by consultancy Yole Group, which was presented by Bosch.
Bosch is a key producer of the tiny sensors, which are crucial in robotics.
At an event in Berlin, Hartung stressed the importance of the components in improving the dexterity of robots.
These sensors determine whether the robot “should tighten its grip or not, whether it is dealing with a sturdy object, or whether it needs to act delicately because it is an egg,” he said.
“Humans have four million touch sensors. If we were to build robots equipped with as many sensors, four years of global sensor production would barely be enough to equip 12,500 robots,” he added.
The focus on automation is also meant to boost the competitiveness of Bosch’s German factories and plug shortages of skilled labour.
Bosch, also known for making a wide range of industrial equipment and household appliances, struck a deal with German robotics firm Neura in January to gather data on factory work.
Under the partnership, several thousand workers in some of Bosch’s 350 facilities worldwide will wear sensor suits to glean training data for Neura robots.
Neura said Wednesday it had raised up to $1.4 billion in fresh fundraising from backers including Bosch, chip giant Nvidia, Amazon and crypto group Tether.
The funding will be used to accelerate Neura’s activities, ranging from the deployment of robots to the rollout of “gyms” for clients to train bots, and the development of the company’s physical AI systems.
The factory floor’s last manual process gets an AI upgrade
Jon Stojan
The floor’s last rule: Hands only
Every autonomous process on a factory floor follows a predetermined path:
The craft behind the calluses
A traditional finishing apprenticeship begins with feel. The early months are less about technique than about calibration: learning how the body interprets pressure and how the same motion produces different results under different conditions. Apprentices learn abrasive selection and progression. They learn how a given resin responds to heat and how ambient humidity affects how a coating lays. That education has genuine value, but it requires four to six months to reach a productive level and years to reach mastery. In a labor market where manufacturing competes against every sector for the same generation of workers, that timeline is increasingly difficult to sustain.
“Surface finishing has always been treated as an art, something you learn through years of practice. But it is physics, and once you model it correctly, you can build systems that learn and adapt in ways that traditional robots can’t,” said Ariyan Kabir, Co-Founder & CEO of GrayMatter Robotics. “The breakthrough for us came when we realized that the skill operators develop over years is really their internalized understanding of physics in action. Encode that physics in software and you can deploy that capability anywhere.”
Autonomous finishing cells change what operator training needs to produce. Workers who once required deep manual craft now need systems fluency, the understanding of what the machine is doing and why, rather than the physical capacity to replicate it.
FAQs
1. Why has surface finishing resisted autonomous solutions longer than other manufacturing processes?
Every other major manufacturing process follows a predetermined path, but surface finishing doesn’t. Geometry variation is the primary factor: no two parts present exactly the same surface, and materials behave differently under heat and pressure. Surface flaws vary part to part, and pre-programmed systems that execute fixed instructions lack the continuous adaptive judgment that finishing requires.
2. How do vision systems enable robots to adapt to different part geometries?
Vision systems allow autonomous finishing cells to read each part’s actual geometry in real time rather than executing a path written to a CAD model. Because part dimensions vary within normal manufacturing tolerances, a system that adapts to what is physically present rather than what the drawing specifies handles geometry variation and surface irregularities without manual reprogramming between parts.
3. What integration challenges should manufacturers expect when adding autonomous finishing to existing production lines?
The most common issues involve floor space allocation, electrical and pneumatic capacity, and workflow sequencing. Autonomous finishing cells are designed as standalone stations that fit into existing layouts without requiring line redesigns. Most facilities run autonomous and manual processes in parallel during an initial validation period before transitioning fully.
Jon Stojan
DIGITAL JOURNAL
June 10, 2026
June 10, 2026
Photo courtesy of Envato.
Opinions expressed by Digital Journal contributors are their own.
Inside an RV plant in northern Indiana, a curved fiberglass countertop that used to take an hour to hand-finish now leaves the line in six minutes. Sanding and polishing a complex surface to a uniform sheen had long resisted autonomous finishing, and for a counterintuitive reason: the variation inherent in geometry and surface condition made it the hardest job on the floor to teach a machine. By 2023, more than four million industrial robots were operating on factory floors worldwide, according to the International Federation of Robotics’ World Robotics report. While CNC machining handled the cuts and robotic assembly handled the bolts, surface finishing held out longer than either of them.
Opinions expressed by Digital Journal contributors are their own.
Inside an RV plant in northern Indiana, a curved fiberglass countertop that used to take an hour to hand-finish now leaves the line in six minutes. Sanding and polishing a complex surface to a uniform sheen had long resisted autonomous finishing, and for a counterintuitive reason: the variation inherent in geometry and surface condition made it the hardest job on the floor to teach a machine. By 2023, more than four million industrial robots were operating on factory floors worldwide, according to the International Federation of Robotics’ World Robotics report. While CNC machining handled the cuts and robotic assembly handled the bolts, surface finishing held out longer than either of them.
The floor’s last rule: Hands only
Every autonomous process on a factory floor follows a predetermined path:
A weld traces the seam an engineer drew
A mill follows the tool path a CAM system generated
Even bin-picking, long the benchmark problem for difficult autonomous applications, became tractable once vision systems learned depth perception
Surface finishing has no such path. Part geometry and material thickness vary. The surface flaws on any given part, such as a small ridge left by a mold or a soft spot in the gel coat, also vary from piece to piece. A skilled finishing operator senses pressure through the wrist, hears the tone of the abrasive shift, watches resin dust change color as it heats, and adjusts continuously. That kind of real-time adaptive judgment can’t be reduced to a routine written months earlier in an engineering office.
GrayMatter Robotics, a Physical AI company building Factory SuperIntelligence (FSI) for manufacturing, deploys autonomous finishing cells designed for exactly that constraint. Where software AI systems learn from internet data, Physical AI systems operate in and learn from the physical world. GrayMatter Robotics’ autonomous finishing cells draw on ATLAS, the company’s proprietary data regime comprising 7 petabytes of real-world surface finishing data accumulated across 30 million square feet, 20-plus industries and 11-plus sensing modalities. That foundation develops Process Intelligence, the learned understanding of how tools, materials and surfaces interact under real manufacturing conditions, enabling the system to adapt in real time to whatever geometry and surface condition the part presents, without pre-programming. The result is closer to what a craftsperson does than to what a conventional robot does.
A mill follows the tool path a CAM system generated
Even bin-picking, long the benchmark problem for difficult autonomous applications, became tractable once vision systems learned depth perception
Surface finishing has no such path. Part geometry and material thickness vary. The surface flaws on any given part, such as a small ridge left by a mold or a soft spot in the gel coat, also vary from piece to piece. A skilled finishing operator senses pressure through the wrist, hears the tone of the abrasive shift, watches resin dust change color as it heats, and adjusts continuously. That kind of real-time adaptive judgment can’t be reduced to a routine written months earlier in an engineering office.
GrayMatter Robotics, a Physical AI company building Factory SuperIntelligence (FSI) for manufacturing, deploys autonomous finishing cells designed for exactly that constraint. Where software AI systems learn from internet data, Physical AI systems operate in and learn from the physical world. GrayMatter Robotics’ autonomous finishing cells draw on ATLAS, the company’s proprietary data regime comprising 7 petabytes of real-world surface finishing data accumulated across 30 million square feet, 20-plus industries and 11-plus sensing modalities. That foundation develops Process Intelligence, the learned understanding of how tools, materials and surfaces interact under real manufacturing conditions, enabling the system to adapt in real time to whatever geometry and surface condition the part presents, without pre-programming. The result is closer to what a craftsperson does than to what a conventional robot does.
The craft behind the calluses
A traditional finishing apprenticeship begins with feel. The early months are less about technique than about calibration: learning how the body interprets pressure and how the same motion produces different results under different conditions. Apprentices learn abrasive selection and progression. They learn how a given resin responds to heat and how ambient humidity affects how a coating lays. That education has genuine value, but it requires four to six months to reach a productive level and years to reach mastery. In a labor market where manufacturing competes against every sector for the same generation of workers, that timeline is increasingly difficult to sustain.
“Surface finishing has always been treated as an art, something you learn through years of practice. But it is physics, and once you model it correctly, you can build systems that learn and adapt in ways that traditional robots can’t,” said Ariyan Kabir, Co-Founder & CEO of GrayMatter Robotics. “The breakthrough for us came when we realized that the skill operators develop over years is really their internalized understanding of physics in action. Encode that physics in software and you can deploy that capability anywhere.”
Autonomous finishing cells change what operator training needs to produce. Workers who once required deep manual craft now need systems fluency, the understanding of what the machine is doing and why, rather than the physical capacity to replicate it.
Five jobs gone, or five jobs changed?
The arithmetic that initially looks like subtraction often resolves differently. When a single autonomous cell handles the work that previously occupied six finishers, the instinct is to read that as five positions eliminated. In practice, facilities that deploy these systems tend to expand rather than contract. The throughput gains that justify autonomous finishing are the same gains that make it possible to take on volumes previously out of reach. Two or three cells running in parallel require upstream support and operational oversight the prior headcount wasn’t providing. The floor is busier, not quieter, and the roles shift accordingly.
A second force is accelerating this transition. CAD and CAM tools have made it cheaper to design parts with curves and undercuts that an engineer would have avoided a decade ago because no one could finish them consistently at volume. Research published in Materials found that conventional CNC machining strategies relying on fixed step sizes are inherently inefficient for surfaces with rapidly varying curvature and that aligning tool paths to local surface geometry reduced form error by 48.4% in a single pass.
Facilities still relying on manual finishing are quietly discovering that the geometry coming out of engineering has outpaced what a human finisher can produce consistently at scale. Geometry-agnostic autonomous systems convert that pressure into a competitive advantage. The complex surface that once represented a bottleneck becomes another part moving through the queue.
The arithmetic that initially looks like subtraction often resolves differently. When a single autonomous cell handles the work that previously occupied six finishers, the instinct is to read that as five positions eliminated. In practice, facilities that deploy these systems tend to expand rather than contract. The throughput gains that justify autonomous finishing are the same gains that make it possible to take on volumes previously out of reach. Two or three cells running in parallel require upstream support and operational oversight the prior headcount wasn’t providing. The floor is busier, not quieter, and the roles shift accordingly.
A second force is accelerating this transition. CAD and CAM tools have made it cheaper to design parts with curves and undercuts that an engineer would have avoided a decade ago because no one could finish them consistently at volume. Research published in Materials found that conventional CNC machining strategies relying on fixed step sizes are inherently inefficient for surfaces with rapidly varying curvature and that aligning tool paths to local surface geometry reduced form error by 48.4% in a single pass.
Facilities still relying on manual finishing are quietly discovering that the geometry coming out of engineering has outpaced what a human finisher can produce consistently at scale. Geometry-agnostic autonomous systems convert that pressure into a competitive advantage. The complex surface that once represented a bottleneck becomes another part moving through the queue.
FAQs
1. Why has surface finishing resisted autonomous solutions longer than other manufacturing processes?
Every other major manufacturing process follows a predetermined path, but surface finishing doesn’t. Geometry variation is the primary factor: no two parts present exactly the same surface, and materials behave differently under heat and pressure. Surface flaws vary part to part, and pre-programmed systems that execute fixed instructions lack the continuous adaptive judgment that finishing requires.
2. How do vision systems enable robots to adapt to different part geometries?
Vision systems allow autonomous finishing cells to read each part’s actual geometry in real time rather than executing a path written to a CAD model. Because part dimensions vary within normal manufacturing tolerances, a system that adapts to what is physically present rather than what the drawing specifies handles geometry variation and surface irregularities without manual reprogramming between parts.
3. What integration challenges should manufacturers expect when adding autonomous finishing to existing production lines?
The most common issues involve floor space allocation, electrical and pneumatic capacity, and workflow sequencing. Autonomous finishing cells are designed as standalone stations that fit into existing layouts without requiring line redesigns. Most facilities run autonomous and manual processes in parallel during an initial validation period before transitioning fully.
Thinking before picking: Smarter harvest robots and the impact on modern farming
Dr. Tim Sandle
Canadian agriculture: early signals of a similar shift
While the specific “harvest-ease” framework is emerging from Japan, Canadian agriculture has already been moving in a similar direction, particularly in greenhouse operations.
In Ontario and British Columbia, greenhouse tomato producers have invested heavily in automation, driven by labour shortages and rising costs. Companies such as Nature Fresh Farms and SunSelect Produce have adopted advanced environmental controls, robotics, and AI-based crop monitoring. Although most harvesting in these facilities is still human-led, there is growing experimentation with robotic assistance systems.
Canadian agricultural technology firms have also begun to focus on decision-support rather than pure automation. For example startups in Ontario have developed vision systems that assess fruit ripeness, orientation, and picking readiness, feeding this information to human workers or semi-automated tools. With a different development, robotics developers in Quebec have explored adaptive gripping systems that adjust force and angle based on fruit characteristics. Furthermore, greenhouse operators are trialling AI-driven crop analytics to prioritise which sections of a crop should be harvested first
These efforts reflect a broader trend: automation in agriculture is shifting away from rigid, pre-programmed actions toward systems that evaluate conditions and adapt in real time.
Human-robot collaboration, not replacement
One of the more practical implications of the research is the model of shared workload it enables. Rather than replacing human labour entirely, the system supports a division of effort. Here, robots handle straightforward, high-probability picks and human workers focus on complex, delicate, or obstructed fruit.
This hybrid model is already emerging in Canadian greenhouses, where labour shortages have made full staffing difficult. In British Columbia’s Lower Mainland, growers have reported challenges in maintaining consistent harvest teams, leading to interest in technologies that can stabilise output rather than eliminate labour altogether. A robot that can filter out easy tasks and leave only the more complex work for humans reshapes productivity. Instead of matching human performance across all tasks, robots can contribute where they are most effective, improving overall efficiency without requiring full autonomy.
A critical next step for systems like Fujinaga’s is moving beyond controlled test environments into real farms, where variability is the norm. Factors such as changing light conditions, plant density, humidity, and unexpected obstructions introduce layers of complexity.
Canadian greenhouse operations, with their controlled yet dynamic environments, may provide an ideal testing ground. These facilities already maintain detailed environmental data and standardised crop layouts, making them suitable for integrating adaptive robotic systems. Field agriculture presents a greater challenge. Outdoor conditions introduce variability that requires even more sophisticated decision-making. However, the principle of assessing task difficulty before execution remains applicable across both settings.
Design implications for agricultural technology
The concept of “harvest-ease estimation” has broader implications for how agricultural robots are designed. This leads to systems built to prioritise tasks based on probability of success, rather than attempting all tasks equally. It is similarly important that robots adjust their approach dynamically, rather than relying on fixed movement patterns. These forms of decision-making models can be combined with crop monitoring systems to optimise harvesting schedules.
The idea of a robot “thinking before acting” may seem incremental, but it marks a shift in how automation is conceptualised in agriculture. Instead of focusing solely on mechanical capability or visual recognition, the emphasis moves toward decision-making under uncertainty. For Canadian farmers, particularly those operating in high-value greenhouse sectors, this aligns with ongoing efforts to modernise production without compromising quality. As automation becomes more adaptive, the barrier to adoption may lower, opening the door to wider deployment across farms of varying sizes.
In practical terms, this approach supports more efficient harvesting, reduced waste, and a clearer path toward collaboration between human workers and machines. For an industry managing labour constraints, cost pressures, and the need for consistent output, the ability to prioritise and adapt may prove as important as the ability to automate itself.
Dr. Tim Sandle
DIGITAL JOURNAL
June 6, 2026
Sprinklers water a lettuce field in California’s Imperial Valley, a vital part of America’s huge agricultural sector, in February 2023 – Copyright AFP FOCKE STRANGMANN
Robotic harvesting has long been positioned as a solution to one of agriculture’s most persistent challenges, which is labour shortages. But while automation has advanced rapidly in areas such as seeding, spraying, and monitoring, harvesting delicate crops has remained a stubborn frontier. Tomatoes exemplify the problem. They grow in dense clusters, ripen unevenly, and are easily damaged, making them difficult targets for machines that rely on simple detection and repetitive motion.
A new research effort from Osaka Metropolitan University suggests a shift in approach. Rather than asking whether a robot can identify and pick a tomato, the system evaluates how easy it will be to harvest each fruit before attempting the task. This change, from detection to decision-making, signals a broader evolution in agricultural robotics, one that aligns more closely with the realities of commercial farming.
Most harvesting robots today are built around visual recognition systems. Cameras and machine learning models identify ripe fruit, then robotic arms attempt to pick it. The limitation is that identification alone does not account for the physical complexity of the plant environment. A tomato may be fully ripe but partially hidden, blocked by stems, or positioned in a way that increases the risk of damage.
Assistant Professor Takuya Fujinaga’s system introduces what can be described as a “harvest probability” model. The robot analyses multiple variables, variations like fruit position, stem orientation, surrounding leaves, and occlusion, before calculating the likelihood of a successful pick. It then selects the approach angle most likely to succeed.
In testing, this method delivered an 81% success rate, with a notable proportion of successful picks coming after the robot adjusted its strategy mid-task. When a frontal approach failed, the system recalculated and attempted a side-angle harvest, demonstrating a level of adaptive behaviour that has been largely absent from earlier systems. For agricultural operators, this represents a move toward machines that behave less like automated tools and more like semi-autonomous workers capable of situational decision-making.
Why this matters for commercial farming:
Harvest efficiency is not simply about speed; it is about yield quality, labour substitution, and operational continuity. Failed picks can damage fruit, slow down processes, and reduce overall productivity. By introducing a system that evaluates how likely a task is to succeed before execution, the research addresses several operational constraints:Reduced damage rates, as difficult picks can be avoided or approached differently
Higher overall throughput, as time is not wasted on low-probability attempts
Improved resource allocation, allowing robots to focus on high-yield tasks
This approach aligns with how human pickers operate. Experienced workers instinctively assess whether a tomato can be picked cleanly and adjust their movements accordingly. Translating this judgement into machine logic is a significant step toward closing the gap between human and robotic performance.
Robotic harvesting has long been positioned as a solution to one of agriculture’s most persistent challenges, which is labour shortages. But while automation has advanced rapidly in areas such as seeding, spraying, and monitoring, harvesting delicate crops has remained a stubborn frontier. Tomatoes exemplify the problem. They grow in dense clusters, ripen unevenly, and are easily damaged, making them difficult targets for machines that rely on simple detection and repetitive motion.
A new research effort from Osaka Metropolitan University suggests a shift in approach. Rather than asking whether a robot can identify and pick a tomato, the system evaluates how easy it will be to harvest each fruit before attempting the task. This change, from detection to decision-making, signals a broader evolution in agricultural robotics, one that aligns more closely with the realities of commercial farming.
Most harvesting robots today are built around visual recognition systems. Cameras and machine learning models identify ripe fruit, then robotic arms attempt to pick it. The limitation is that identification alone does not account for the physical complexity of the plant environment. A tomato may be fully ripe but partially hidden, blocked by stems, or positioned in a way that increases the risk of damage.
Assistant Professor Takuya Fujinaga’s system introduces what can be described as a “harvest probability” model. The robot analyses multiple variables, variations like fruit position, stem orientation, surrounding leaves, and occlusion, before calculating the likelihood of a successful pick. It then selects the approach angle most likely to succeed.
In testing, this method delivered an 81% success rate, with a notable proportion of successful picks coming after the robot adjusted its strategy mid-task. When a frontal approach failed, the system recalculated and attempted a side-angle harvest, demonstrating a level of adaptive behaviour that has been largely absent from earlier systems. For agricultural operators, this represents a move toward machines that behave less like automated tools and more like semi-autonomous workers capable of situational decision-making.
Why this matters for commercial farming:
Harvest efficiency is not simply about speed; it is about yield quality, labour substitution, and operational continuity. Failed picks can damage fruit, slow down processes, and reduce overall productivity. By introducing a system that evaluates how likely a task is to succeed before execution, the research addresses several operational constraints:Reduced damage rates, as difficult picks can be avoided or approached differently
Higher overall throughput, as time is not wasted on low-probability attempts
Improved resource allocation, allowing robots to focus on high-yield tasks
This approach aligns with how human pickers operate. Experienced workers instinctively assess whether a tomato can be picked cleanly and adjust their movements accordingly. Translating this judgement into machine logic is a significant step toward closing the gap between human and robotic performance.
Canadian agriculture: early signals of a similar shift
While the specific “harvest-ease” framework is emerging from Japan, Canadian agriculture has already been moving in a similar direction, particularly in greenhouse operations.
In Ontario and British Columbia, greenhouse tomato producers have invested heavily in automation, driven by labour shortages and rising costs. Companies such as Nature Fresh Farms and SunSelect Produce have adopted advanced environmental controls, robotics, and AI-based crop monitoring. Although most harvesting in these facilities is still human-led, there is growing experimentation with robotic assistance systems.
Canadian agricultural technology firms have also begun to focus on decision-support rather than pure automation. For example startups in Ontario have developed vision systems that assess fruit ripeness, orientation, and picking readiness, feeding this information to human workers or semi-automated tools. With a different development, robotics developers in Quebec have explored adaptive gripping systems that adjust force and angle based on fruit characteristics. Furthermore, greenhouse operators are trialling AI-driven crop analytics to prioritise which sections of a crop should be harvested first
These efforts reflect a broader trend: automation in agriculture is shifting away from rigid, pre-programmed actions toward systems that evaluate conditions and adapt in real time.
Human-robot collaboration, not replacement
One of the more practical implications of the research is the model of shared workload it enables. Rather than replacing human labour entirely, the system supports a division of effort. Here, robots handle straightforward, high-probability picks and human workers focus on complex, delicate, or obstructed fruit.
This hybrid model is already emerging in Canadian greenhouses, where labour shortages have made full staffing difficult. In British Columbia’s Lower Mainland, growers have reported challenges in maintaining consistent harvest teams, leading to interest in technologies that can stabilise output rather than eliminate labour altogether. A robot that can filter out easy tasks and leave only the more complex work for humans reshapes productivity. Instead of matching human performance across all tasks, robots can contribute where they are most effective, improving overall efficiency without requiring full autonomy.
A critical next step for systems like Fujinaga’s is moving beyond controlled test environments into real farms, where variability is the norm. Factors such as changing light conditions, plant density, humidity, and unexpected obstructions introduce layers of complexity.
Canadian greenhouse operations, with their controlled yet dynamic environments, may provide an ideal testing ground. These facilities already maintain detailed environmental data and standardised crop layouts, making them suitable for integrating adaptive robotic systems. Field agriculture presents a greater challenge. Outdoor conditions introduce variability that requires even more sophisticated decision-making. However, the principle of assessing task difficulty before execution remains applicable across both settings.
Design implications for agricultural technology
The concept of “harvest-ease estimation” has broader implications for how agricultural robots are designed. This leads to systems built to prioritise tasks based on probability of success, rather than attempting all tasks equally. It is similarly important that robots adjust their approach dynamically, rather than relying on fixed movement patterns. These forms of decision-making models can be combined with crop monitoring systems to optimise harvesting schedules.
The idea of a robot “thinking before acting” may seem incremental, but it marks a shift in how automation is conceptualised in agriculture. Instead of focusing solely on mechanical capability or visual recognition, the emphasis moves toward decision-making under uncertainty. For Canadian farmers, particularly those operating in high-value greenhouse sectors, this aligns with ongoing efforts to modernise production without compromising quality. As automation becomes more adaptive, the barrier to adoption may lower, opening the door to wider deployment across farms of varying sizes.
In practical terms, this approach supports more efficient harvesting, reduced waste, and a clearer path toward collaboration between human workers and machines. For an industry managing labour constraints, cost pressures, and the need for consistent output, the ability to prioritise and adapt may prove as important as the ability to automate itself.
No comments:
Post a Comment