The authors develop an offline estimator of state empowerment.
Offline experience can reveal where agents have greater future control
The authors estimate empowerment from unlabeled trajectories and use it to guide online exploration in sparse-reward tasks.
Top University
Ishir Garg · Vivek Myers · Qiyang Li · Catherine Ji · Benjamin Eysenbach · Sergey Levine
UC Berkeley · Princeton University
Research Digest··2 min read
Garg et al.
§