Robots Train Themselves
Robots are now training themselves in the real world. 🤖 NVIDIA's ENPIRE framework lets AI agents autonomously run physical experiments with a 99% success rate. Tap to see how it works!
What the video says
A massive shift is coming to physical robotics as machines begin to train themselves in the real world. Researchers at NVIDIA have developed a software framework called Empyre, which establishes a closed-loop system for robots to autonomously try, fail, and learn from their tasks.
Using a physical feedback routine across 4 core modules, Empyre allows coding agents to run autonomous experimentation loops directly on physical hardware, In testing, frontier coding agents like GPT-5.5 and Opus 4.7 autonomously developed policies that achieved a 99% success rate on complex physical tasks. This watershed moment in automation is supported by other massive infrastructure projects such as Tencent running its Argus software on a cluster of over 10,000 GPUs to detect slow hardware failures.
As these systems scale, some observers are warning of a highly consequential future. Author Fernando Berretti suggests that, in the long run, everyone made of flesh and blood will be disempowered and replaced by machines.
