SFTW Spotlight: Farmwise - Reinventing Agriculture Equipment

Programming Note: SFTW will be back to regular programming from the coming weekend. You can subscribe to the free newsletter here.

Welcome back to another edition of Startup Spotlight with SFTW. This edition features Farmwise, an AI and robotics company providing weeding services using a computer vision based system using mechanical blades.

I had the privilege to connect with Tjarko Leifer, CEO of Farmwise, and Greg Chiocco, VP of Product Management of Farmwise a few weeks ago, including a lab and a field visit. The write up is based on my interactions with the Farmwise team, and products.

Full disclosure: I am NOT an investor in Farmwise, NOR do I have any personal financial stake in the company. This is NOT a sponsored post.

I. Labor is the big challenge

A growing labor shortage and rising wages are a source of increasing concern for the global agriculture industry, and especially here in California. Labor typically accounts for more than 50% of total production costs, and most growers expect labor costs to grow at 10-30% over the next three to five years!

Growers have been increasingly looking to technology solutions to help alleviate some of the challenges due to labor. For example,

UC Davis publishes production costs for different crops in different parts of California every year. According to their study for romaine lettuce, hand weeding costs were $ 284 per acre in 2023.

Screenshot from UC Davis Study (page 14) on production costs for romaine hearts lettuce in California.

There is a clear customer problem to be solved, and many companies are stepping into the space to address this problem in different ways.

II. Technology convergence opens the door for viable solutions

The launch of Uber in 2009 was enabled by a convergence of many technologies coming together and being more generally available. GPS had become prevalent, iPhone had launched in 2007, mobile commerce was taking off due to better payment rails, and the economic anxieties of the global financial crisis all came together to help propel ridesharing as a business.

There is a similar convergence happening within the agriculture sector.

Vision systems and cameras are dropping in price, sophisticated AI modeling capabilities and processing units capable of doing computations on the edge, mechanical robotic arms and blades working in tandem with software are now available.

The technology convergence combined with rising labor costs for agriculture work, and policy challenges due to the existing political environment is creating an opening for different approaches to solve the labor problem using technology.

We see a similar phenomenon in commodity row crops, which is being driven by high input costs for chemicals, and issues like pesticide resistance are leading to solutions like selective spot spraying for crop threats like weeds.

Farmwise is doing the same, but starting with specialty crops. Farmwise is building products to help specialty crop growers reduce their high labor cost, improve operational efficiencies, and address labor training issues.

Here is a video to give you a sense of Farmwise’s flagship product, the Vulcan robotic weeding implement.

III. Farmwise’s mission and vision

Farmwise has a big bold mission and vision.

Enable growers and ag equipment OEMs to reimagine every farming task with machinery that can perceive, understand, and act at the plant level in real time

Farmwise believes,

Farming can be done at the plant level, at scale, thereby harnessing natural variability of biological systems

Farmwise plans to digitize specialty crop agriculture at the plant level, which can accelerate insights. They plan to do it by harnessing the natural variability to improve the sustainability of farming practices.

According to Farmwise, their solution

Decreases dependency on labor with intra and inter-row weeding with sub-inch precision no more weeding crew
Turns every operator into best operator with an open architecture, micro adjustments from cab, and remote software support
Unlocks operational efficiency by being compatible with existing tractor fleet, lightweight, with ease of setup, repair, and maintenance
Gets the job done fast as it can work with 3 beds at once, at 2 acres per hour whatever the field conditions

and provides the following benefits

Address labor shortages by automating repetitive, physically challenging tasks
Improve quality of operations by monitoring performance and automatically adjust machine configurations
Reduce crop protection and fertilizer inputs by applying the right amount of the right chemistry at the right place and right time, or maybe not use chemistry at all.

Farmwise believes they can do this by reinventing agriculture equipment.

IV. Reinvent agriculture equipment

US agriculture has followed a one-size-fits-all approach enabled by mechanization, use of chemicals, and biotechnology to reduce the agriculture workforce to 1-2% of the total labor force compared to 60% about 100 years ago.

The technology convergence I mentioned above, is presenting an opportunity to revisit this one size one-size-fits all approach.

Can we reinvent how agriculture equipment is designed, developed, and deployed?

Farmwise definitely thinks so.

If you go back to Farmwise’s mission, they want to enable growers and ag OEMs to reimagine every farming task and act at the plant level in real time.

Farmwise does not ask growers to change their farming practices in terms of how they plant, what they plant, and when they plant different specialty crops. Farmwise’s system (called Vulcan) can be attached to the back of a regular tractor, and so does not require a huge behavior change from the grower.

The grower can continue to plant their crops with the same row spacing as before, and plant the same way as before. Farmwise’s system is modular and can be configured to do weeding on a different number of rows at the same time as necessary.

Farmwise’s Vulcan robotic weeding implement at the back of a regular tractor somewhere in California (Image provided by Farmwise team)

Vulcan is a sophisticated system with cameras, a strobe lighting system to get clear images for the AI models, edge computing capabilities powered by artificial intelligence models, and a mechanical blade system which can be used for different purposes like weeding, and cultivation.

The choice of using a mechanical weeding approach with a blade moving through the soil is a conscious one. Farmwise decided to focus on the software connected to a mechanical actuation system (which is a known entity with many different approaches already available from other industries like manufacturing, assembly lines, etc.), rather than build a new mode of actuation, for example lasers.

The modular design and inbuilt flexibility into Vulcan does not force a grower to use a one size fits all approach and take advantage of natural variability inherent within agriculture.

The Vulcan system helps automate manual agriculture labor like weeding. Due to its mechanical weeding approach, Vulcan can substitute some of the chemistry or biotech which in turn has substituted much of the labor over the last hundred years.

Image of a weedy field before and after a Vulcan pass. As you can see in the right hand side, most of the weeds have been taken out by Vulcan. (Image provided by Farmwise)

V. Vulcan robotic weeding implement

Vulcan provides a few settings which provide flexibility for your weeding and cultivation operation.

Number of lines (you can decide how many rows you want to operate on)
Width of the blade opening (this can be adjusted based on row spacing of your crop)
Margin for how close you get to the plant (if you have many weeds close to the plants, you can use this setting)
Blade speed (how quickly do the blades open and close)

Farmwise has invested in building a strong MLOps system (MLOps = Machine Learning Operations). MLOps is a paradigm to deploy and maintain machine learning models in production reliably and efficiently. The word is a compound of "machine learning" and the continuous development practice of DevOps in the software field.

Farmwise has used their strong MLOps system, combined with a large data set of 300,000 labeled images (with 200 million total images) to build out an Intelligent Plant System (IPS) for weeding. IPS has the following features:

One central RGB camera. Farmwise plans to transition to a triple stack camera with 2 stereo cameras on the outside
The lighting package has 1100 watts of light which are strobed for a well-lit scene to get high quality images for the AI algorithms. It removes the need for a heavy and cumbersome hood to block out shadows or varying lighting conditions.
The IPS strobes the lights quickly, and generates only 2-3 Watts of heat, which can be pulled out through a heat sink without the need for fans.
The IPS includes NVIDIA edge processors for performance at speeds of 2-3 miles per hour. Farmwise AI models can run at 20 miles per hour, but soil dynamics limit the speed at which it can travel. If a different actuation mechanism is used (for example, spraying), the same IPS system can be used for faster operational speeds.

The setup for Vulcan is straightforward

High quality weeding starts in the cab with an intuitive in-cab interface for configuration adjustments and customization.
Set-up is similar to a traditional cultivator. Reconfigurations can be made in under 60 minutes. Each self-contained weeder module makes automatic micro adjustments based on bed terrain.

Foundation models

One of the challenges with agriculture is the infinite variety of contexts. This makes it difficult to build solutions which can scale quickly, as it is oftentimes hard to go from one use to the next. For example, if you have solved a problem with strawberries, it is not always easy to solve a similar problem in raspberries.

Foundation models are a newer concept (2021) within the world of artificial intelligence. A foundation model is a machine learning or deep learning model, trained on broad data and can be applied across a wide range of use cases.

Farmwise uses foundation models to train and deploy models for new crops efficiently, based on the models built for existing crops. They can enable new use cases within the existing diversity of conditions.

For example, when one of Farmwise’s customers wanted to do weeding in collard greens, it was relatively easy for Farmwise to add a collard green model, even though they had not built a collard green model before.

A practical example was provided in a LinkedIn post by Greg Chiocco, VP of Product at Farmwise.

Why didn’t the FarmWise Vulcan robotic weeder kill all those white flowers in the lettuce field? Because we trained our foundation model to recognize it as Alyssum, a beneficial insectary plant that attracts good insects to organic fields.

Foundation models are important as you can get decent performance out of the box for a new use case. As more data comes in for new use cases and crop types, it improves existing models. For example, getting additional data on field shapes, shape of furrows and beds, in a new crop will make the model for an existing crop better, as it gets better at dealing with different types of field shapes, shape of furrows and beds etc.

I have not seen many companies in agriculture use foundation models successfully to enable new use cases. (The common foundation models are ChatGPT and Gemini from Google)

The use of a strong ML Ops system, combined with foundation models will give the right tools to Farmwise to scale into new use cases along crop types, operations, and regions.

Farmwise’s decision to focus on hydraulic based mechanical actuation, gives it more flexibility rather than a new mode like lasers, which is limited to only a “kill” operation.

Farmwise has provided some guidance on their roadmap, and the different operations they plan to support in the future.

VI. Farmwise Business model

How does Farmwise plan to make money?

Most robotics companies in the specialty crop space have moved from robotics as a service to a grower purchase and operated model. Outright ownership by the grower is a good risk mitigation strategy for the grower, given the importance of timeliness of operations like weeding, etc.

If a grower owns the equipment, they are guaranteed to have the equipment available when they need it.

This increases the CapEx requirements for the grower, many growers in California and other western states are willing to spend the money due to labor challenges, as they see a tangible ROI from the use of solutions like Farmwise. This is an example from Farmwise.

(Growers) found it to be quite versatile to send Vulcan to a field at a moment's notice, which is far easier than scheduling a crew of 25 via a labor contractor. In the end, the customer cited Vulcan's versatility, productivity and weeding quality as key components in their purchase decision. Given their weeding costs the ROI was quite favorable compared to other technology purchases they have made on the farm.

Farmwise plans to go down a partnership driven approach with agriculture OEMs, which is enabled due to the versatile and modular nature of IPS, and foundation models. It can give Farmwise access to the OEM’s distribution, and support network. Farmwise is still thinking through the OEM partnership model, but my guess is it might involve a few components.

One time integration and development investment in a new use case
Variable hardware cost of putting the camera, lighting, and compute package
Ongoing cost of training model drift, performance improvement, and support requirements.

The price point for Vulcan is on the higher side today. Farmwise’s strategy is to start at the high end, and then ride down the cost curve as more use cases are enabled, and the BoM costs continue to go down. We have seen this example in the automotive space, where new technology is available to high end brands. The technology and products enabled by the technology can then ride down the cost curve to mass markets.

VII. Farmwise team

Farmwise has a strong and experienced leadership team.

Farmwise is led by Tjarko Liefer, with extensive experience in building and scaling startups. Tjarko was the head of strategy and operations at The Climate Corporation when it was acquired by Monsanto for close to a billion dollars. Tjarko founded a company called Wellio, which was acquired by Kraft Heinz. Wellio built an AI platform for meal planning and grocery shopping based on a dataset of 12+ million recipes and 700k grocery items. He is on the board of Arable.

Farmwise’s product team is led by Greg Chiocco. Greg is one of the smartest people working in AgTech. He has extensive experience in building hardware and software products, taking them to market and scaling them. He has experience working at Trimble, The Climate Corporation (Monsanto and Bayer Crop Science), Zoox (self-driving company acquired by Amazon), Granular, and Mineral.

The rest of the Farmwise team includes experts in hardware, robotics, artificial intelligence, electrical and mechanical systems etc.

VIII. Conclusions

Labor problems are real in agriculture, especially in specialty crops, and they are going to get worse. The convergence of various technology advancements have created the right conditions for advanced robotic solutions for operations like weeding to be commercially viable.

The future success of Farmwise will depend on a few factors, given an acquisition by an existing OEM or a agriculture retailer is one of the most viable exit options.

Can they continue to attract the right type of talent for their commercial and technology teams?
Can they find the right pricing and business model for working with industry partners?
Can they tap into the right sales and distribution channels to scale?
Can they continue to evolve their technology to make it easier and easier to add and enable new use cases?

What do you think?

If you like “Software is feeding the world”, please share with a friend.

If you don’t mind answering 3 questions anonymously (2 are optional), I would love to get your feedback.

This newsletter is a product of Metal Dog Labs, LLC

Unsubscribe | Update your profile | Metal Dog Labs LLC, Dublin, CA 94568