TL;DR
By applying abstract harmonic analysis to discrete morphological symmetries in robots (like identical legs on a quadruped), this paper decomposes complex robotic motions into orthogonal, independent linear subsystems. Leveraging this property—termed Dynamics Harmonic Analysis (DHA)—the authors introduce the equivariant Dynamics Autoencoder (eDAE), which learns globally linear Koopman models with dramatically higher sample efficiency, robustness, and interpretability.
What Problem Does This Paper Address?
Traditional robotic control relies on Lagrangian rigid-body dynamics, which are nonlinear and necessitate iterative, local linearizations (e.g., in iLQR or Differential Dynamic Programming). These local approximations can trap controllers in local minima and complicate estimation.
Koopman operator theory provides an appealing alternative: it maps nonlinear dynamics into an infinite-dimensional space where the evolution is strictly linear. However, learning finite-dimensional data-driven approximations of the Koopman operator is notoriously challenging due to high sample complexity, sensitivity to noise, and the curse of dimensionality.
Crucially, existing Koopman modeling methods ignore morphological symmetries—the physical duplications of limbs and kinematic chains ubiquitous in robotics.
Methods
The paper bridges group representation theory, harmonic analysis, and deep data-driven Koopman modeling:
- Isotypic Decomposition: Using the Peter-Weyl theorem and group representation theory, the state space $\mathcal{X}$ of a symmetric system is decomposed into orthogonal isotypic subspaces ($\mathcal{X} = \bigoplus_i \mathcal{X}_i$). Each subspace corresponds to an irreducible representation (irrep) of the robot’s discrete symmetry group $G$.
- Dynamics Harmonic Analysis (DHA): The authors prove that any $G$-equivariant linear model (such as a Koopman operator) written in the isotypic basis decouples into independent, block-diagonal subsystems ($K_{\Delta t} = \bigoplus_i K_{\Delta t, i}$). By Schur’s Lemma, each block is further constrained into repeated scalar multiples of identity maps, drastically reducing free parameters.
- Equivariant Dynamics Autoencoder (eDAE): Extending the standard Dynamics Autoencoder (DAE), eDAE enforces:
- A $G$-equivariant neural encoder $f_\phi$, guaranteeing that the latent space inherits the symmetry structure.
- A block-diagonal, equivariant transition matrix $K_{\Delta t}$ aligned with the isotypic subspaces, eliminating spurious correlations across independent modes.
x(t) ──[ G-equivariant Encoder ]──> z(t) ──[ Block-Diagonal K_Δt ]──> z(t+Δt)
│
x(t+Δt) <───────[ Decoder ]──────────────────────────────────────────────────┘
Datasets
- Synthetic Constrained Linear Systems: Linearly constrained dynamical systems with white-noise perturbations subjected to cyclic groups ($C_5$ and $C_{10}$). Evaluated across varying state dimensions, latent dimensions, and noise levels.
- Quadruped Closed-Loop Locomotion Dataset: Trajectories of a Unitree Mini-Cheetah quadruped running a model predictive controller (MPC) with a periodic trotting gait on uneven terrain. The numerical state space is 37-dimensional (joint positions, joint velocities, base height, orientation quaternion, and tracking velocity errors). Its discrete symmetry group is $G = K_4 \times C_2$ (order 8).
Key Insights and Ideas
- Normal Modes for Robotics: Just as molecules have normal vibrational modes, symmetric robots have fundamental, symmetry-constrained synergistic motions. Complex gaits decompose cleanly into projections within these lower-dimensional subspaces.
- Sparse Energy Distribution: Real-world motions like trotting or jumping concentrate kinetic energy within only one or two dominant isotypic subspaces; other subspaces are only activated transiently during disturbances.
- Geometric Regularization Beats Data Augmentation: Standard data augmentation forces networks to learn equivariance through brute force. Structuring the latent space with exact group equivariance and block-diagonality inherently prevents the model from fitting non-physical cross-talk between independent dynamic modes.
Key Results
- Sample Efficiency & Generalization: On the Mini-Cheetah, eDAE achieved substantially lower test Mean Squared Error (MSE) than both vanilla DAE and data-augmented DAE (
DAEaug), requiring significantly fewer trajectories to reach low prediction errors. - Full Symmetry Exploitation: Models utilizing the complete symmetry group $G = K_4 \times C_2$ consistently outperformed models utilizing only the subgroup $K_4$, validating that modeling all discrete symmetries directly enhances predictive accuracy.
- Robustness: Across synthetic benchmarks, eDAE demonstrated flat, stable prediction errors under increasing state dimensionality, wide ranges of latent space sizes, and high noise standard deviations ($\sigma$), whereas baseline DAE degraded rapidly.
Limitations and Future Work
- Discrete Symmetries Only: The framework specifically addresses finite/discrete morphological symmetry groups; continuous symmetries (e.g., full $SO(3)$ or planar translations) are not handled directly by this formulation.
- Known Symmetries Required: The group representations $\rho(g)$ on the physical coordinate space must be known a priori.
- Parallel Optimization in Trajectory Planning: The authors note that applying DHA to decouple local linearizations in classic algorithms (like iLQR and DDP) could allow parallelized computation of sub-models—a promising avenue left for future research.
Summary updated