RL Research - What Tasks Would You Want to Use?

In RL research, I feel like that's one of the major bottlenecks. It's the lack of data sets and benchmarks where you can really explore all of these different directions in generalization. So we had to sort of take an existing benchmark and artificially change the simulation to look different. But it's still pretty limited to how much diversity you can get from that. What tasks would you want to use in order to explore continual learning in an environment like this?