Welcome, sci-fi realists…

San Francisco AI startup Dimensional is paying people $20 an hour to walk robot dogs. No leashes, treats or little plastic bags required. So what exactly is the point?

Turns out, these walks are less about exercise and more about education. Dimensional’s robot dogs use cameras and sensors to collect real-world data as they move through the city, helping train the AI systems that power them. And unlike walking actual dogs, the job requires no experience.

But before you quit your day job to become a professional robot dog walker, there are a few catches. Shifts run four to six hours, and if your mechanical companion breaks down, you're responsible for hauling all 33 pounds of it back to the office.

Still, $20 an hour to help train the next generation of robots isn't a bad gig. Just remember: If your dog suddenly refuses to walk, you can't bribe it with a treat. You're carrying it home.

Anyway, here's what you need to know about robotics right now:

AGIBOT’s new AI model gives humanoid robots a serious performance boost

Humanoid robots are getting better at doing actual work, and AGIBOT has some impressive numbers to back that up. In a recent trial, the company's G2 humanoid completed tasks with an 85.7% success rate using its new Go-3 Preview AI model. Without it? Just 29%. That's a pretty substantial leap from getting most things wrong to getting most things right.

The test took place in a convenience store, where AGIBOT put its G2 through a series of tasks before and after introducing Go-3. The secret sauce is a combination of human demonstration footage and robot-generated data, which helps humanoids learn how to handle unfamiliar environments rather than simply repeat rehearsed movements.

AGIBOT also fine-tuned the model to tackle multi-step tasks, including building LEGO structures, and its Genie Evolver training system takes things further, allowing robots to practice tasks and evaluate their own performance. Think of it as homework for humanoids, except they're learning how to manipulate physical objects instead of writing essays.

Why it matters: Sure, a robot that makes your bed correctly 86% of the time sounds appealing, even if you'd still have to straighten the sheets occasionally. But the bigger opportunity is industrial. Warehouses and other workplaces need robots that can reliably handle different objects and adapt when conditions change. Go-3's results suggest progress toward that goal, although a successful convenience-store trial doesn't automatically translate to warehouse reliability.

AGIBOT also plans to open-source the full model by the end of 2026, potentially giving other robotics developers a head start. For now, the machines are getting smarter. Whether they're ready to clock in for a full shift is another question.

Robots in Action

Remember when we reported on China experimenting with humanoid robots as traffic cops? Well, the robots have officially reported for duty. At the Hangzhou Marathon, humanoids from DEEP Robotics made their debut as traffic police assistants, directing runners along the course and helping keep the event moving smoothly.

They're not exactly writing speeding tickets yet, but it's another step toward putting humanoids to work in public spaces. And for once, the robots aren't the ones being asked to run.

Check them out below:

Around the Horn

Want to get more out of whatever AI tools you’re using? The Neuron Academy offers self-paced courses to help professionals use AI in their work. You’ll find courses on the best ways to research, create content, and write better prompts, plus how to build custom workflows, agents, and apps. The academy can even recommend courses based on your goals. Check out the scope of courses we offer, and get a leg up on the humanoids.

Robots Behaving Badly

Apparently, robots are learning bad driving habits from humans. On October 26, a Starship Technologies delivery robot ran a red light in Tallinn, Estonia, forcing a city bus to brake sharply and reportedly injuring several passengers. And then things got even stranger.

According to local transit officials, the six-wheeled delivery bot left the scene before police arrived. Starship disputes the hit-and-run interpretation, saying the robot contacted its remote support team, which moved it to a safe spot nearby.

The company is investigating the incident and plans software improvements to prevent similar crashes. It has also rerouted its robots away from the intersection.

The good news? Nobody needs to teach robots how to drive drunk. The bad news? They apparently don't need alcohol to drive badly.

Industry Signals

P.S. If you haven’t already… follow us on Youtube!

Not subscribed? Subscribe here.