Introduction
Imitation learning typically assumes a direct mapping between an expert's actions and a learner's capabilities. However, when the expert and learner possess different physical structures—such as different joint counts or limb lengths—the "correspondence problem" arises. Traditional solutions like kinesthetic teaching or manual teleoperation are resource-intensive and lack scalability. This paper proposes a mathematical framework to automate similarity mapping between dissimilar robotic embodiments, allowing a learner to translate an expert's spatial configuration into its own unique state space.
Methodology
The authors introduce a metric-based framework that avoids manual joint-to-joint mapping by focusing on global spatial similarity.
Embodiment State and Distance Metric
The state of a robot is represented by frames (homogeneous transformation matrices in $SE(3)$) and twists (velocities in $se(3)$) for each link. A composite distance function $d(s, \hat{s})$ quantifies the difference between the expert state $s$ and learner state $\hat{s}$ using three components: translational distance, rotational distance (using a cosine-based measure for stability), and velocity distance.
Link Correspondence
To handle varying link counts, a correspondence matrix $W$ is employed. This can be a static pre-calculated mapping or a "soft" state-dependent mapping using a softmax function to identify nearest-neighbor links in real-time. This "softmax trick" ensures smooth transitions and prevents jerky movements during imitation.
Learning Architectures
The framework supports two learning modes:
- Static Pose Imitation: A neural network learns to map expert joint angles to learner joint angles by minimizing the distance metric as a loss function.
- Dynamic Movement Imitation: The distance metric serves as a negative reward signal within a Proximal Policy Optimization (PPO) reinforcement learning loop, enabling the agent to learn trajectories without prior knowledge of its own dynamics.
Experiments and Results
The framework was validated using 2D planar manipulators and high-fidelity simulations of Franka Emika Panda arms in Gazebo.
Pose and Trajectory Generalization
Neural networks successfully generalized pose translations from a 7-DOF expert to a 4-DOF learner. In dynamic tasks, the PPO agent learned to follow complex trajectories. Notably, training on a diverse set of trajectories significantly reduced generalization error compared to single-trajectory training. Even when restricted to only 3 DOFs, the learner successfully minimized the distance metric by aligning its available links with the general spatial "silhouette" of the 7-DOF expert.
Conclusion
This research demonstrates that a well-defined spatial distance metric can bridge the gap between dissimilar robotic bodies. While the optimization landscape can occasionally suffer from local minima, the combination of rotational alignment and static correspondence provides a robust foundation for automated imitation learning. Future work may extend this "free topology" approach to more diverse robotic forms, such as swarms or non-anthropomorphic systems.
Summary updated