DayDreamer: World Models for Physical Robot Learning
Applies Dreamer to 4 robots to learn online and directly in the real world, without any simulators, which establishes a strong baseline.
Dreamer algorithm can facilitate fast learning on physical robots using a world model. This can help robots solve tasks in complex environments, learn from small amounts of interaction, and reduce the amount of trial and error needed in the real environment.
Joint Generator-Ranker Learning for Natural Language Generation
Achieves new SotA performance on five public benchmarks covering three popular generation tasks: summarization, question generation, and response generation.
JGR is a novel joint training algorithm that integrates the generator and the ranker in a single framework, effectively harmonizing their learning and enhancing their quality. It can improve generation quality on various text generation tasks and surpasses existing methods on four public datasets across three common generation scenarios.