From Farm to Factory: How Open-Source Edge AI Is Cleaning Up China’s Tomato Supply Chain
Chaihuo Mobile Creative Vehicle is a roving AI lab and makerspace launched by Chaihuo Maker Space. Over the course of 200+ days, the Chaihuo MCV is on the journey of travelling deep into China’s interior — traversing three major terrain steps, six climate zones, and more than ten distinct landform types — bringing edge computing, digital fabrication, and open-source hardware workshops to underserved communities. Along the way, the project stress-tests technology in demanding environments, co-develops solutions with local communities, and uncovers real-world use cases where open hardware can make a meaningful difference.
It was during one such stop — in Xinjiang, home to the lion’s share of China’s processing tomato output — that the MCV team sat down with a partner they already knew well, and brought solutions to their chronic problems.
Every harvest season in Xinjiang, thousands of truckloads of tomatoes pour out of mechanized fields and into processing plants. These tomatoes aren’t destined just for your salad bowl. They’ll be transformed into tomato paste, ketchup, and sauce, bound for shelves around the world. China supplies roughly a quarter of the world’s processing tomatoes, with Xinjiang alone responsible for three-quarters of that national output, more than one million mu (~67,000 hectares) under cultivation.
But hidden inside those truckloads — sometimes quite literally — is a problem that better farming practices alone cannot solve: rocks, mud, plastic, weeds, and other debris mixed in with the fruit.
The Hidden Cost in Every Truckload
At one of our partners’ companies — an agricultural technology firm working with more than 20 tomato processing factories — the team identified a pattern that was quietly eating into their clients’ margins. When farmers deliver tomatoes by the truckload (typically 30 tonnes per load), some deliveries arrive containing up to 17% impurities by weight. Stones, clods of earth, and other foreign matter get mixed in during mechanical harvesting — and in some cases, are deliberately added to inflate the load’s weight.
For processing plants, this isn’t merely an inconvenience — it’s a serious financial problem. Factories pay farmers by the weight of each delivery. If a 30-tonne truck carries 5 tonnes of rocks and soil, the plant is effectively paying top money for rubble. Making matters worse, the only corrective measure available — manual sorting — introduces its own inefficiencies: worker fatigue and human inconsistency mean that even the sorting process can’t reliably catch what the scales can’t distinguish.
Our partner had already developed sophisticated software platforms for crop monitoring, harvest scheduling, and remote-sensing analysis. But the hardware layer — the physical devices needed to capture what was actually on the conveyor belt, where the vehicles were, and how the equipment was performing — was where critical gaps remained.
Two Problems That Off-the-Shelf Hardware Couldn’t Solve
When our partner came to us, they had two interconnected challenges that no single vendor in the market could address:
1. Impurity detection on the conveyor belt. The partner’s vision was ambitious: deploy AI-powered cameras at multiple checkpoints across the sorting line — on the harvester’s extended boom arm, at the primary inspection platform, along the elevator conveyors, and beside the manual sorting stations.
2. Predictive equipment maintenance. (This is an emerging area of collaboration that we’re exploring together.) If hard debris enters high-speed processing equipment — crushers, pulpers, centrifuges — it can cause abnormal vibrations that damage internal components. Detecting these anomalies early could shift maintenance from “fix it when it breaks” to “fix it before it breaks.” The team is evaluating vibration sensor modules equipped with an AI training platform capable of frequency spectrum analysis, allowing the system to learn each machine’s normal vibration signature and flag deviations in real time.
The partner explored commercial solutions, but ran into the same wall each time. Industrial cameras offered standard video feeds but locked users out of any low-level customization. And there was simply no product on the market that brought together video input, GPS data, and local edge compute in a single open, modifiable package.
Why Open Hardware Changed the Equation
Before this project even began, our partner was already a long-time Seeed customer — they had been running SenseCAP weather stations across their operations for years. That existing familiarity with our hardware is what brought them to us when the challenge grew more complex.
This is where our approach diverges from the typical vendor-client relationship. Rather than building a custom turnkey solution — which would have been expensive, slow, and locked to our engineering bandwidth — we delivered standardized, fully open-source hardware and let our partner’s own engineering team do the development.
Here’s what they built with it:
reCamera for AI-powered impurity recognition. Impurity detection was the partner’s most pressing priority. To address this, they decided to deploy reCamera 2002 units across the four checkpoint positions on the processing line. reCamera runs Linux on a RISC-V AI SoC with 1 TOPS @INT8 computing power and comes with YOLO11 built in — giving the partner’s engineers a fully open platform to develop custom models and deploy them directly on the device. The team fine-tuned two YOLO11 variants for this application: the lighter YOLO11s model for on-site sorting at 640px resolution, and a higher-fidelity YOLO11m model (1024×1024) for offline quality auditing. The entire AI pipeline — from data collection to model deployment — was built and validated by the partner’s engineers themselves, using reCamera’s open architecture.
Robotic arm sorting — an emerging frontier. Some factories in the partner’s network have already begun experimenting with robotic arms to physically remove debris from the sorting line. In our MCV, a reBot Arm is also equipped. The partner showed strong interest and is eager to take it further: pairing AI-powered visual recognition with mechanical actuators to replace the most labor-intensive step in the process. Both sides are still in early-stage validation on this front.
Edge computing on the RK3588 platform. For on-vehicle compute, the team evaluated several compute tiers and identified the reComputer RK3588 AI Box as the preferred option, for its balance of AI performance, power consumption, and cost. The plan is to purchase one or two units for initial testing and validation before moving to full deployment.
The Open-Source Advantage: Why This Model Scales
The Open-Source Advantage: Why This Model Scales
None of this required custom development on our end. The partner’s engineers built everything themselves — the detection models, the tracking integration, the full AI pipeline — using hardware whose schematics, firmware, and board-level documentation are entirely open. We shipped standard products; they built the expertise.
That’s also why the solution can scale. A successful pilot at one factory can be replicated across the 20+ facilities in their network without coming back to us for customization.
At Seeed, we’ve always believed that the most meaningful technology isn’t the kind that gets built for communities — it’s the kind that communities can build for themselves. Open-source hardware is how we act on that belief: not by solving problems on behalf of people, but by putting the tools in their hands and getting out of the way.
The tomato project is one example of what that looks like in practice. When local engineers can train their own models, build their own trackers, and scale their own solutions — without depending on a vendor for every iteration — that’s SDG 9 (Industry, Innovation and Infrastructure) in action. And when smarter sorting and dispatch means less food lost to debris and spoilage across an entire regional industry, that’s a real contribution to SDG 2 (Zero Hunger) that no single product could have delivered on its own.
Beyond Tomatoes: One Stop Among Many
The tomato project is one stop on a much longer journey. Since leaving Shenzhen, the Chaihuo MCV has passed through 21 provinces and 36 cities — connecting with 22 makers and touching 3 industries along the way.
At Aksu Apple Orchards, a local agricultural-mechanization engineer is using XIAO microcontrollers, webcams, and motors to prototype a robotic arm for apple harvesting. Same playbook: we provide the hardware, they bring the domain expertise.
Out on the Tagong Grassland in western Sichuan, the team is working with a local herder to establish the first maker learning space in the region, starting with yak tracking and expanding into workshops and educational programs across the plateau.
👇Follow the full route — and every co-creation along the way.
About Author
