Verifying Source Authenticity of Successful Robotic Demonstrations

Determine a reliable verification procedure that, given a robotic manipulation demonstration that appears successful, ascertains whether the behavior was generated by an autonomous policy rather than by human teleoperation, thereby resolving ambiguity in trajectory provenance (Source Authenticity).

Background

The paper identifies two key ambiguities undermining trustworthy evaluation of robotic manipulation: execution quality and source authenticity. Even when a trajectory visually appears successful, existing benchmarks provide no definitive mechanism to verify whether the behavior was produced autonomously or via hidden human teleoperation.

This provenance uncertainty prevents fair comparison of policies and leaves evaluations vulnerable to manipulation. The authors position source discrimination as a necessary component of a trustworthy evaluation standard and propose dataset design and modeling approaches aimed at addressing this gap.