Welcome to the era of embodied AI: Introducing AA-Robotics v0.1

Artificial Analysis releases a new benchmark which ranks the ability of LLMs to perform robotics tasks. Only a few models are currently released on the benchmark. Several models are able to complete most tasks on the benchmark although the speed is painfully slow. GPT-6.1 Sol seems best overall when time to complete and accuracy are both considered. I'd be interested to see some older and open models tested on this benchmark too. I'm curious whether any of them could do anything on this benchmark at all. I personally think that the sudden ability of LLMs to do robotics tasks reasonably well is a huge and underrated opportunity for continued scaling and post-training paths. Given that several models can complete most of these simple tasks (albeit slowly), there seems to be a clear path to scaling robots through LLMs. While we're scraping the bottom of the barrel for easily reward-verifiable tasks for LLMs to be trained on in many text-native fields, there are innumerable easily reward-verifiable tasks that can be constructed for an LLM with vision and audio as the input and robot arm controls (or humanoid robot controls) as the output.