Meet Lynx M20: a new set of eyes for agriculture. With our client, we’re exploring how computer vision can help robots count cattle. Farmers can fully delegate herd monitoring and fence inspection, helping them see more and spot problems sooner. Is Black Mirror here already?

We spend a lot of the robot-teaching time in two places: real video, and simulation. What you are seeing is the real action, then the same task in simulation on our robot. A person gets through the motion much faster than we can teleoperate it, which is why we keep filming real work instead of living on robot hours. The long part is everything between the two shots. We take the human motion from the video, put it on our joints, and run it in sim until the robot is actually doing it. A motion that looks right on the person does not land on our arm until we have worked it over. Sim is where that happens, because we can repeat it all day before we put it on hardware. We are done with a clip when the arm in sim can do what the person did. Filming it is just the start.

Can a robot become an animal supervisor? We’re about to find out.

A factory robot works because the product never changes. An orchard breaks that assumption completely. Different size, different angle, different height, half hidden behind leaves. That's why real-world footage is the bottleneck.

There's a way to get paid for work you're already doing. It's called FRAME Capture. You open the app, pick a task from the list, film yourself doing it, and submit. If it's filmed correctly, credits go straight to your account. All you need is your phone. 👉 framerobotics.ai/capture

Before you teach your robot something, check if someone already did it. Install a skill instead of building one. FRAME Skill Store is the place for you!

We're taking a closer look at JEV. The split we already wanted looks like this. A language model as the brain. A decision model as the gate. The llm reads the task, the demo, the scene. Jev does not write a plan and it does not move joints. It scores the next commit. Grasp or abort. Retry or hold. Is this contact stable enough to keep going. That stack is technically real. Slow layer proposes. Fast layer votes. Controller runs the arm. Jev is the vote in a form software can consume. A choice, a score, a yes or no, with a probability. Cheap enough for the decision loop. Not the 1khz loop. That still belongs to the hardware. Jev cannot see a camera. Your code has to turn the scene into state first. It cannot be the safety layer. If confidence is low, you stop in code. There is still a lot of work. Test it on real ticks. Measure whether 80% confident is actually right 80% of the time on your robot. Perception, rates, fallbacks, what happens when the scene dump is wrong. None of that is done. JEV-class models in robotics are still promising. Worth the research time. Because the gate between "the model said pick" and "the gripper closed" finally has a shape you can afford to call.