Introduction
In imitation learning, the "correspondence problem" arises when a robot must replicate a demonstrator's actions despite having a fundamentally different physical structure. For instance, how does a four-legged Sony AIBO robot interpret a two-legged human's gesture? Historically, measuring the success of such imitation has been largely subjective. This research introduces a formal mathematical metric to quantify pose similarity across diverse embodiments, moving beyond qualitative assessment to a rigorous scalar measure.
Methodology: The k-Tree Abstraction
The core of the proposed solution lies in representing agents as open kinematic chains. By abstracting an agent's body into a k-tree, where each physical link is represented as a unique edge, the researchers can focus on the underlying geometry of a pose rather than the specific "skin" or scale of the robot. This abstraction allows a human limb to be mathematically compared to a robot dog's hind limb.
To compare different bodies, the algorithm utilizes homeomorphic pose classes. This allows the partitioning of edges into disjoint sets to find a "common minimal homeomorph"—a simplified skeleton shared by both agents. The dissimilarity between two poses is then calculated as the summed squared difference of the instantaneous slopes of these paired edges. The result is a pseudometric that satisfies the triangle inequality, providing stability for machine learning applications.
Experimental Validation and Results
The metric was tested using a dataset comprising 48 poses from three distinct agents: a human subject, a Sony AIBO robot, and a simulated three-segment dolphin. The researchers captured 16 poses for each agent to perform all-pairs distance testing.
The results demonstrated that the metric successfully identified "closest neighbors" across embodiments. For example, a human standing on one leg with arms outstretched was mathematically mapped to the most geometrically similar pose in the AIBO's repertoire. These quantitative matches aligned consistently with qualitative human judgment, confirming the metric's effectiveness in solving the correspondence problem for static poses.
Conclusion and Future Directions
By providing a scalar measure of dissimilarity, this work enables robots to mathematically select the best action to imitate a human demonstration. While currently limited to static poses, future research aims to extend this framework to time-extended actions. By treating sequences of poses as continuous curves in pose space, the authors hope to facilitate the imitation of fluid motion, such as walking or complex manipulation tasks.
Summary updated