Auditing Learned Command Adapters for Locomotion Policies

This research introduces an audit framework to determine if learned adapters for frozen locomotion policies provide genuine performance gains. Testing on a Go2 robot revealed that while some headroom existed, actual recovered allocation gain was only 0.55%, suggesting many adapters may be unnecessary.