Low Cost Robots

11 Jun 2025 by Adam Allevato

Vassar Robotics’ Hacker News announcement of their new robot arm was poorly timed. At least, it was poorly timed with respect to optimizing my bank account balance. Not one day after I coughed up 300ish dollars for the electronics for my own SO-101 arm, including an extra ~50% for import fees, I learned about Vassar’s Navrim robot, which offers the same functionality, but at a significant discount. We’ve had robot arms, cheap servos, and 3D printers for a while, so how did we end up here in 2025? In short, we need more arms to collect demonstration data.

Transformers have completely changed how AI and robotics mesh together. Prior to “the ChatGPT moment”, techniques like CNNs and graph neural networks were used to wholly replace parts of the traditional sense-think-act robotics loop. But using transformers, which can incorporate data from multiple sensors more seamlessly, startups and R&D divisions have cranked out example after example of networks that replace most or all of a robot’s brain.

Training transformers requires a lot of data, and the requirements are even larger when the transformer is multimodal. This is not a new revelation. Physical Intelligence, 1X, Diligent Robotics (my current employer), and many others have recognized that the industry is collectively facing a huge data shortage in our quest to train generalist robotics policies. There is no equivalent to “a text corpus of most human knowledge”, aka “the internet”, that can be used to train robots. A quick way to try to address this shortfall is to make lots of robots that can in turn be used to collect lots of data. And one of the best ways to make lots of robots is to make them cheap - very cheap.

I’m still trying to figure out where and when these new “extreme-economy arms” make a company profitable. In order for this new wave of low cost robots to find a real use case, they will have to solve either the payload problem or the accuracy problem. These are effectively toys, like the Armatron that I had as a kid, and the policies that run on them are toy policies. That can turn into something useful in the long run, but the policies will always be limited by the arm. If payload is too low, then the policy will not lift useful items like cast iron pans or blankets. If accuracy is too low, then the policy will not be able to do dextrous tasks like assembling electronics or slicing a bell pepper. The envelope of achievable tasks for these devices is bounded by low payload on one end and low accuracy on the other, and I’m not sure if there are any useful tasks inside those bounds.

For now, it looks like if my robots are going to clean up after me, I’m going to have to buy a full set of lightweight plastic dishes (and cookware).