Warm-start new agents with teacher-student training
Opened this issue · 0 comments
dtch1997 commented
Currently to iterate on experiments when testing new sensor designs, we continually train new agents from scratch. However, we may be able to greatly reduce training time by training a basic blind walking policy on flat ground and using DAgger or other imitation learning techniques to transfer the basic walking skills to the new agent. This may reduce experiment runtimes by a lot potentially enabling faster iteration of settings.