We open-source all data corresponding to the 80-task and 30-task datasets used in our multi-task experiments. They can be downloaded below. The two datasets contain 545M and 345M transitions, respectively.

We release two multi-task datasets with data from 80 and 30 tasks, respectively. The datasets are summarized below.

These datasets are obtained from the replay buffers of 240 single-task TD-MPC2 agents, and thus contain a wide variety of behaviors ranging from random to expert policies. Multi-task agents trained on the above datasets, as well as checkpoints for each of the single-task agents, are available here.

We list all 80 tasks in the dataset below, along with TD-MPC2 policy visualizations and summary statistics.