Can we learn agents that can autonomously explore complex environments and achieve commanded goals?.Here, we introduce TLDR that trains such agents in complex environments, including Ultra-large AntMaze! . Project Page: Paper:

7/ Please check out our paper and website for more information!.work with.@kwanyoung_park_ .@YoungwoonLee . Project Page: Paper: Code:

6/ TLDR also shows interesting emergent behaviors, like windmilling Ant and quadrupedal Humanoid.

5/ TLDR shows impressive goal-reaching behaviors! For example, it shows superior goal-reaching ability and state coverage in complex environments like AntMaze and Humanoid.

4/ Our method, TLDR, integrates temporal distances into Go-Explore, such that the agent can explore farther goals and learn better at reaching diverse goals!

3/ To address the challenges, we propose using temporal distances (i.e., minimum environment steps between the states) for unsupervised GCRL. First, we learn representations that encode these temporal distances.

2/ Unsupervised RL can learn useful behaviors without supervision (e.g., rewards). In particular, unsupervised goal-conditioned RL can autonomously learn to reach diverse goals. Key challenges are how to 1) select goals for exploration and 2) effectively learn to reach the goals.