Bonsai Robotics ships Bonsai World to train farm robots before they arrive

Bonsai Robotics said Friday it launched Bonsai World, a simulation and world-model tool that turns satellite imagery of farms and mines into 3D environments so autonomous machines can train on conditions they have not yet physically encountered.

Physical AI dies in the gap between the acre you already drove and the one you have not. Bonsai World bets that satellite-to-sim training — grounded in 50M+ real samples — can shrink bring-up time for farm and mine robots before the first wheel hits dirt.

On Friday, 2 October 2026, Bonsai Robotics introduced Bonsai World. The EIN Presswire page is titled “Bonsai Robotics Unveils Bonsai World to Accelerate Physical AI Across Rugged Environments.” The page says the news was provided by Bonsai Robotics and stamps October 02, 2026, 13:00 GMT, which is 9:00 a.m. Eastern. The dateline is San Jose, California. Bonsai World is a simulation and world-model application. A simulation, here, is a computer copy of a place a machine can practice in. A world model is the software that builds that copy, so the machine can train before it drives the real site. Physical AI, in this release, means software that has to run on a machine in the real world, not only answer a question on a screen. The release says the tool lets the company generate environments and conditions machines have not yet physically encountered, to speed the development and deployment of physical AI in rugged, unstructured environments. Unstructured means a place that is not a neat factory aisle: dirt, weather, and things that move. The company calls itself the leader in AI-first autonomy software for the rugged world. Leader is its phrase. Those lines are the wire’s.

Where the new tool sits. Bonsai World is the newest capability inside Bonsai Intelligence. The release calls Bonsai Intelligence the environmental intelligence layer that powers autonomy on the company’s Amiga platform and on retrofitted OEM equipment. Autonomy means the machine drives and works without a person steering every move. OEM is the original equipment manufacturer, the company that built the machine. A retrofit puts Bonsai’s software on equipment that already exists, rather than only on a machine Bonsai built. Amiga is Bonsai’s own vehicle platform. The release does not describe the Amiga’s size, and it does not print a price. Those lines are Bonsai’s.

The field record the models start from. Bonsai said its foundation and world models are trained on more than 50 million real-world samples collected across more than one million acres. The release calls that an industry-leading dataset. Industry-leading is the company’s phrase. The acres span crops, terrain, weather, lighting, machines, and jobs. A sample, here, is one recorded piece of what a machine saw or did in the field. More than 50 million is the company’s count of those pieces. More than one million acres is the company’s count of the land they came from. An acre is a little smaller than an American football field, so a million acres is a wide farming region, not one orchard. The release says Bonsai World extends that intelligence from what the fleet has already experienced to environments it has never encountered. Those figures are Bonsai’s. The page does not print a customer count beside them.

How a satellite picture becomes a practice world. Bonsai World starts with satellite imagery of a farm, a mine site, or another rugged environment and turns that flat map into a structured 3D simulation. Flat, here, means the overhead picture. Three-dimensional means a scene with depth, so a machine can move through it. In that simulation, autonomous machines can recreate real paths and interact with the environment. The system can generate photorealistic ground-level views and introduce conditions such as dust, debris, animals, vehicles, and changing terrain. Photorealistic means the view is meant to look like a photograph taken at ground level. The release says this lets Bonsai train and evaluate autonomy before equipment is deployed on site, and that it reduces how much field data the company has to collect, and how many test-and-fix rounds it has to run, when it enters new crops, machines, jobs, and operating environments. Deployed means the machine has been sent to the site to work. Those lines are the release’s. The page does not print a measured count of hours saved.

The two loops the release describes. Bonsai says its data advantage compounds with every machine deployed and every job completed. Bonsai World multiplies that loop by generating new training data on demand, while real deployments keep expanding and grounding the dataset. On demand means a new practice scene is made when it is needed. Grounding, in that sentence, means the real trips keep the simulated worlds tied to what the machines have actually seen. Together, the release says, the real loop and the synthetic loop make a richer base for rugged physical AI and improve model performance and dependability over time. Synthetic means made in the simulation rather than recorded in a field. Improving performance is the company’s claim. The page does not print a before-and-after score.

What the release says an operator gets. Bonsai World is supposed to improve deployment readiness before a machine arrives, reducing or eliminating bring-up time, field tuning, and repeated test cycles. Bring-up is the work of getting a machine ready on a new site. Field tuning is adjusting it once it is there. The result, in the company’s words, is a faster path to productive work, with the autonomy already trained on the terrain, the conditions, and the edge cases the machine is likely to meet. An edge case is an unusual situation that can still stop a machine: an animal in the row, a washout, dust thick enough to hide the crop. Reducing or eliminating is Bonsai’s claim. The page does not print a clock time for a named farm.

Tyler Niday, chief executive and co-founder, is quoted on the release. “Every environment we operate on expands what the system understands,” he said. “We started in some of the toughest operating conditions we could find—unreliable GPS, limited connectivity and near-zero visibility in dust—because we knew that if our models could handle those environments, they could generalize to many others. Bonsai World lets us compound that experience—to take what we’ve learned in the field, generate the conditions we haven’t seen yet, and prepare the next machine before it ever gets there. That’s how we move from solving autonomy one deployment at a time to building intelligence that can scale across the rugged world.” GPS is the satellite signal a machine uses to know where it is. Unreliable GPS means that signal drops or wanders. Near-zero visibility in dust means the cameras cannot see through the cloud a machine kicks up. Generalize means a model trained in one place still works in another. That quotation is his, in the release.

The computers behind the simulation, as the release states them. Bonsai World combines Google Cloud and NVIDIA accelerated computing with Bonsai’s years of field data. Google’s Gemini VLM interprets the satellite imagery and builds a structured map of the environment. A VLM is a vision-language model: it looks at a picture and turns what it sees into a description a computer can use. Gemini is Google’s model. The release says the Bonsai world model is then post-trained using more than 50 million real-world samples, so the model is grounded in how rugged sites actually look. Post-trained means extra training on top of a model that already exists, using those field samples. Training runs on Google A2 virtual machines with NVIDIA A100 GPUs. A virtual machine is a computer rented in a cloud, not a box in the farm shop. A GPU is a graphics processor, the chip these machines use to train a model and to draw a scene. Generating the simulations, and running the model once it is trained, uses Google G4 virtual machines with NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. Inference is that second step: the model producing a result, rather than the earlier step of training it. The release says the combination makes simulations that stay consistent in time and in shape, and that reflect the conditions its machines actually meet. Consistent in time means the scene does not jump to a different place from one moment to the next. Those product names, and that claim, are the release’s.

Les Karpas, Inception Global Head of Physical AI at NVIDIA, is quoted on the release. He said rugged and unstructured environments are the ultimate test for physical AI, because machines have to adapt to terrain, weather, crops, and crop health that change from place to place, plant to plant, and season to season. He said that, building on NVIDIA Cosmos and using NVIDIA accelerated computing, Bonsai World turns field experience into realistic simulations that help prepare autonomous machines for new jobs and environments before deployment. Cosmos is NVIDIA’s name for the simulation software he says this builds on. That quotation is his, in the release. The page does not describe a separate NVIDIA product launch, and it does not say NVIDIA bought a stake in Bonsai.

Darren Mowry, vice president of Global Startups and Investor Ecosystem at Google, is quoted on the release. He said the Bonsai team is taking an innovative approach to physical AI, applying multimodal models and building new world models for hard problems in agriculture. A multimodal model takes more than one kind of input, such as a picture and words. He said Google Cloud’s AI stack will support Bonsai, including Gemini vision models and the training of Bonsai’s own world models, so the company can orchestrate connected robotic fleets in rugged environments. Orchestrate means coordinate a group of machines. Will support is his sentence about Google Cloud’s tools. His title on the page includes the word investor. The release does not say Google took an ownership stake, and it does not print a funding round or a price.

Who the about box says the company is. Bonsai Robotics calls itself a full-stack physical AI company building the intelligence layer for autonomous machines in the rugged, off-road world. Full-stack, here, means the company says it builds both the software and the machines it runs on. Bonsai Intelligence powers vehicles on the Amiga platform and on retrofitted OEM equipment. Bonsai Pilot lets operators plan and orchestrate those fleets. The platform is deployed in specialty-crop agriculture in the United States and Australia and is expanding into other rugged industries, including mining and defense. Specialty crops are fruits, vegetables, and similar crops, not the big commodity grains. Deployed is the company’s account of where the platform already runs. Expanding is its account of where it is going. The page does not name a mine, a defense customer, or a farm. Media inquiries go to Linda McNair at Bonsai Robotics, +1 831-420-7949, linda.mcnair@bonsairobotics.ai. The page points readers to bonsairobotics.ai.

The picture is a lime-and-earth desk graphic. A yellow bar splits a lime panel from a dark green grid. The lime panel reads BONSAI WORLD, then satellite to 3D sim, 50M+ field samples, train before deploy, Google Cloud + NVIDIA, and Amiga + OEM retrofit. The dark panel reads Physical AI for rugged environments, then dust, debris, and animals, then terrain, weather, and crops, then new simulation and world model. A yellow frame holds a photograph of a gray machine in leafy crops, with a sensor frame over the body and an orange bar on the front. The lime panel is what names that machine class as Amiga. The photograph does not print its own caption. The frame does not print a calendar date. It is not a satellite map, and it is not a photograph of a mine.

In plain terms, Bonsai Robotics said on Friday, from San Jose, that Bonsai World is a new piece of Bonsai Intelligence. It starts from a satellite picture of a farm or a mine and builds a 3D practice world, including dust, debris, animals, vehicles, and changing ground, so an Amiga or a retrofitted machine can train before it arrives. The company says the models behind that world were trained on more than 50 million real samples from more than one million acres. Google’s Gemini reads the satellite picture. Training runs on Google cloud machines with NVIDIA A100 GPUs. The simulations are generated on Google cloud machines with NVIDIA RTX PRO 6000 Blackwell GPUs. The chief executive says the company learned in bad GPS, thin connectivity, and dust, and that Bonsai World is how that experience gets applied to conditions the fleet has not seen. The platform, the company says, already runs in specialty-crop fields in the United States and Australia and is expanding toward mining and defense. The release does not print a price or a customer count, and it does not say NVIDIA or Google bought a stake.