Link preview
How to make toys that your robots will play with for hours
It's underappreciated how close to perfect the performance of a robot needs to be to be profitable, and getting there takes an enormous amount of experimentation across data, hardware, and machine learning. In CV or LLMs, the same test set can be used forever. However in robotics, each test needs to be manually reset and evaluated for success. This does not scale, especially when success is measured as the difference between 98% and 99% success. Here's what that scaling problem costs in practice. Measuring a policy at 90%+ level with any confidence takes 40-50 rollouts per checkpoint ( reddit.com · reddit.com
It's underappreciated how close to perfect the performance of a robot needs to be to be profitable, and getting there takes an enormous amount of experimentation across data, hardware, and machine learning. In CV or LLMs, the same test set can be used forever. However in robotics, each test needs to be manually reset and evaluated for success. This does not scale, especially when success is measured as the difference between 98% and 99% success. Here's what that scaling problem costs in practice. Measuring a policy at 90%+ level with any confidence takes 40-50 rollouts per checkpoint (
Comments