Part 4/8:
One of the most impressive feats is the development of a vast benchmark dataset, consisting of over 100,000 natural language tasks and 60 realistic simulated homes. This dataset serves as a training ground for robots equipped with advanced reasoning capabilities. The Partner framework enabled an 8.6 times increase in training speed, thanks to the innovative application of smaller language models for reasoning. This impressive advancement doesn't come at the cost of accuracy, allowing robots to comprehend and follow instructions effectively.