Introduction
Most biological and robotic systems possess inherent morphological symmetries, such as the bilateral symmetry of a humanoid or the radial symmetry of a multi-legged robot. These symmetries imply that the physical laws governing the "left side" of a system are identical to those of the "right side." However, standard data-driven models, such as standard Multi-Layer Perceptrons (MLPs) or Convolutional Neural Networks (CNNs), are typically "symmetry-blind." They must learn these fundamental physical relationships from scratch, requiring extensive datasets to recognize that a left-foot contact event is functionally identical to a right-foot contact event.
While data augmentation can partially address this, it is computationally expensive and does not provide mathematical guarantees of symmetry. This paper introduces a formal group-theoretic framework to identify and embed these discrete morphological symmetries (DMS) directly into neural network architectures as an inductive bias, resulting in improved sample efficiency and generalization.
Methodology
The authors bridge the gap between classical mechanics and deep learning through several key technical contributions:
- Formal Identification of DMS: Within the framework of Lagrangian mechanics, a DMS is defined as a transformation that leaves the system's kinetic energy and generalized mass matrix equivariant. This allows for the rigorous mapping of physical symmetries to mathematical groups, such as the reflection group $C_2$ or the Klein four-group $K_4$.
- Scalable Equivariant Neural Networks: Building on the Equivariant Multi-Layer Perceptron (EMLP) framework, the authors optimize the architecture for robotics. By recognizing that robotic symmetries are primarily permutations of sensors and joints, they compute network constraints in linear time $O(mn)$, avoiding the polynomial complexity of previous numerical solvers.
- Optimal Initialization: The authors derived a new initialization scheme tailored for equivariant layers. Standard methods like Kaiming initialization do not account for weight-sharing constraints, which can lead to vanishing or exploding gradients in symmetric architectures.
Key Insights: Symmetries as Permutations
A central insight of this work is that discrete robot symmetries can be treated as re-orderings (permutations) of joints and sensors. This transformation turns a complex linear algebra problem into a simple "orbit" search on a graph, making it feasible to apply equivariance to high-dimensional inputs like depth maps or long proprioceptive time-histories.
By hard-coding these symmetries, the number of trainable parameters is reduced by a factor of approximately $1/|\mathcal{G}|$, where $|\mathcal{G}|$ is the order of the symmetry group. For a quadruped with four-way symmetry ($K_4$), the model requires only ~25% of the parameters of a standard MLP to achieve equivalent representational power.
Experimental Results
The framework was evaluated on two primary tasks: Center of Mass (CoM) momentum estimation for the Atlas humanoid and Solo quadruped, and real-world contact detection for the Mini-Cheetah quadruped.
- Sample Efficiency: In the Solo quadruped experiments, equivariant models (E-MLP) achieved lower error with only 10,000 samples than standard MLPs achieved with 60,000 samples.
- Parameter Reduction: For the Mini-Cheetah contact detection task, the equivariant CNN reduced the memory footprint of basis matrices from several Terabytes (if stored densely) to just a few Megabytes using sparse representations.
- Generalization: Equivariant models demonstrated superior robustness when tested on "out-of-distribution" gaits and terrains not encountered during training, as the structural bias constrained the model to physically plausible symmetric patterns.
Conclusion and Future Work
While the current framework assumes "perfect" symmetry, real-world hardware often possesses small asymmetries due to manufacturing tolerances or wear. Future research will focus on approximate equivariance, allowing models to learn the nearly symmetric nature of physical systems while still benefiting from the structural efficiency of the group-theoretic approach. Additionally, extending the framework from discrete groups to continuous symmetries (e.g., arbitrary rotations for robotic arms) remains an open area for development.
Summary updated