Beyond Topology: A Morphological Symmetry Graph Representation for Locomotion Policy Learning
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: I'm Rosa.
Dev: And I'm Dev, here with Taro.
Rosa: We're looking at the paper: "Beyond Topology: A Morphological Symmetry Graph Representation for Locomotion Policy Learning".
Dev: Here's the gist of the paper. Reinforcement learning has enabled impressive locomotion skills on articulated robots, but common policy representations remain only weakly aligned with robot physics Starting from the robot’s topological graph, our representation augments each observation and action space with the permutation and sign transformations induced by morphological symmetry This yields a symmetry-equivariant graph actor and a symmetry-invariant graph critic, enforcing the desired policy and value constraints by construction rather than through reward shaping or data augmentation Experiments show improved symmetry generalization, robustness, sample efficiency, and model efficiency over topology- and symmetry-aware baselines The Core Representation We introduce a morphological symmetry graph representation for locomotion policy learning We represent an articulated robot as a graph and complete this graph with morphological symmetry actions on observations, privileged observations, and actions, including the permutations and sign changes induced by sagittal reflection Built on this representation, MS-PPO uses a morphological-symmetry-equivariant graph actor and a morphologicalsymmetry-invariant graph critic The goal is not to define a new RL optimizer, but to define a physically aligned function space in which policy and value functions respect both the robot’s kinematic structure and its morphological symmetry How the Training Signal is Built The representation involves augmenting the topological graph with morphological symmetry actions on observations, privileged observations, and actions The actor identity is policy equivariance, meaning mirrored observations produce mirrored actions The critic identity is value invariance, meaning morphologically symmetric states have the same value This is achieved by defining the actor as policy equivariance, where mirrored observations produce mirrored actions <ref:2512.00727#pg2>, and the critic as value invariance, meaning morphologically symmetric states have the same value <ref:2512.00727#pg5> The encoder h(·) maps physical robot signals into graph features that transform only by graph reindexing under G, as in Eq.
Rosa: (3) Where the Gains Come From Experiments show improved symmetry generalization, robustness, sample efficiency, and model efficiency over topology- and symmetry-aware baselines MS-PPO consistently outperforms strong morphology- and symmetry-aware baselines, achieving the best joint-failure tolerance, the lowest sim-to-real velocity MSE on G1 with only 34.6% of standard PPO’s parameters <ref:2512.00727#pg7> The results show that jointly encoding kinematic topology and morphological symmetry is an effective representation design for locomotion policy learning on articulated robots MS-PPO uses only 34.6% of the parameters of PPO and PPO-Aug and 69.3% of the parameters of PPO-EMLP, suggesting that jointly encoding kinematic connectivity and morphological symmetry improves model efficiency without sacrificing task-level performance <ref:2512.00727#pg7> Evaluation Across Embodiments We evaluate MS-PPO on the Unitree Go2 quadruped and G1 humanoid robots MS-PPO achieves the best velocity tracking among all deployable methods, while PPO and MI-PPO fail to produce stable hardware locomotion <ref:2512.00727#pg8> These results support our central claim that effective morphology-aware policy learning requires encoding symmetry structure consistently within the robot’s topological graph representation <ref:2512.00727#pg8> MS-PPO improves symmetry generalization, robustness to asymmetric joint failures, sample efficiency, model efficiency, and zeroshot sim-to-real deployment relative to topology-aware and symmetry-aware baselines Limitations Our current validation is limited to bilateral C2 symmetry on one quadruped and one humanoid <ref:2512.00727#pg9> We have not yet tested richer or approximate symmetry structures, such as rotational Cn robots, reflective K4 quadrupeds, radially arranged multi-arm systems, or symmetric manipulators <ref:2512.00727#pg10> Another limitation is that the current experiments instantiate the representation separately on each robot, rather than demonstrating policy transfer across embodiments with different node features, action spaces, and symmetry orbits <ref:2512.00727#pg10> Future work will address these limitations by automating symmetry graph construction and symmetry action deriving <ref:2512.
Paper summary: Dev: So, wrapping up this look at Beyond Topology: A Morphological Symmetry Graph Representation for Locomotion Policy Learning. The core message here is that the way we represent a robot's body matters more than just connecting the joints; we need to explicitly encode how physical properties change when you reflect or mirror parts of that body.
Rosa: Yeah, it moves beyond just using the robot's topology as a starting point and instead augments it with these symmetry actions to create a representation that is both policy-equivariant and value-invariant <ref:two thousand five hundred twelve point zero zero seven two seven#pg3. This means the AI learns policies that respect the robot's physical structure, not just its connectivity, which helps it be more robust in the real world.
Taro: For someone looking at this from a deployment standpoint, what this really means is that we are building a system where the underlying mathematical structure of the robot’s shape is baked into the learning process itself rather than being an afterthought. It’s about making sure that when you deploy it on something new, or something slightly different, it has a much better chance of working because it already understands the physics of symmetry.
Dev: Exactly. And the results are compelling, showing improved efficiency and performance compared to existing topology-aware and symmetry-aware methods <ref:two thousand five hundred twelve point zero zero seven two seven#pg7. It shows that jointly encoding kinematic topology and morphological symmetry is an effective design choice for locomotion policy learning on articulated robots.
Rosa: So, while they’ve made a strong case for this representation, the authors are being clear about where they still need to go, specifically needing to test it across more complex or approximate symmetry structures <ref:two thousand five hundred twelve point zero zero seven two seven#pg10. It's a solid step forward in alignment between policy and robot physics.
Conclusion: Rosa: So, this paper introduces something called Beyond Topology: A Morphological Symmetry Graph Representation for Locomotion Policy Learning. It’s about taking the basic graph structure of a robot and adding information about its shape's symmetry—how it looks when you flip or mirror parts of it.
Dev: Right, so instead of just focusing on how the joints connect, they are augmenting the observation and action spaces with those permutation and sign transformations caused by morphological symmetry. That means the AI learns policies that respect the robot’s actual physical shape, not just its wiring diagram.
Taro: From my side, what this implies is that we’re building a system where the physical rules of how a robot moves are baked directly into how it learns to walk or run in real-time. If you can teach it symmetry equivariance and invariance at the representation level, it should generalize better when things get messy outside the lab.
Rosa: Exactly. They show that using this setup, they get a graph actor that is equivariant—meaning if you mirror the robot's observation, the action mirrors too—and a graph critic that is invariant—meaning symmetric states have the same value. It enforces these physical constraints by construction instead of relying on tweaking reward functions or adding extra data.
Dev: And the numbers back up what they’re saying, though they keep things focused. They found better symmetry generalization, more robustness when joints fail in asymmetric ways, and improved sample efficiency compared to existing topology-aware and symmetry-aware methods. It even uses much less of the policy parameters than some other complex setups.
Taro: That efficiency is key for deployment because it means we don't need massive amounts of data just to get a stable gait. But I have to point out their limitations, they only tested this on very specific, bilateral C2 symmetry in one quadruped and one humanoid robot right now.
Rosa: True. And another thing they noted is that they built this representation separately for each robot instead of showing how the policy transfers when you switch between robots with different features or action spaces. That's a big gap for deployment across different platforms, right?
Sizhe Wei, Xulin Chen, Fengze Xie, Garrett Ethan Katz, Zhenyu Gan, Lu Gan
Georgia Institute of Technology · California Institute of Technology
cs.RO
Submitted: 2025-11-30
Updated: 2026-11-09
Comments: Project page: https://lunarlab-gatech.github.io/MS-PPO/
Code: https://github.com/Teddy-Liao/walk-these-ways-go2
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 93/100
The gist: Reinforcement learning has enabled impressive locomotion skills on articulated robots, but common policy representations remain only weakly aligned with robot physics Starting from the robot’s…
Key concepts
- Morphological Symmetry Graph Representation
- This method models a robot as a graph where nodes are joints and edges are connections. It enhances this graph by including actions and observations that account for the robot's inherent morphological symmetry, such as reflections or sign changes. This ensures the learned policy respects the physical structure of the robot.
- Symmetry-Equivariant Graph Actor
- The actor in this system is designed to be symmetry-equivariant. This means if you input an observation that has been mirrored or transformed by a morphological symmetry operation, the resulting action will also be mirrored or transformed in a corresponding way. It enforces physical consistency during policy learning.
- Symmetry-Invariant Graph Critic
- The critic is designed to be symmetry-invariant. This means that if two different states are morphologically symmetric—meaning one can be transformed into the other by a symmetry operation—the critic assigns them the same value. This helps the system learn a more consistent understanding of robot states.
- Policy Equivariance vs. Value Invariance
- The actor enforces policy equivariance, ensuring mirrored inputs yield mirrored outputs. The critic enforces value invariance, ensuring that morphologically symmetric states share the same predicted value. These properties are built into the representation to guide the learning process physically.
Terminology
Summary
Reinforcement learning has enabled impressive locomotion skills on articulated robots, but common policy representations remain only weakly aligned with robot physics Starting from the robot’s topological graph, our representation augments each observation and action space with the permutation and sign transformations induced by morphological symmetry This yields a symmetry-equivariant graph actor and a symmetry-invariant graph critic, enforcing the desired policy and value constraints by construction rather than through reward shaping or data augmentation Experiments show improved symmetry generalization, robustness, sample efficiency, and model efficiency over topology- and symmetry-aware baselines
The Core Representation
We introduce a morphological symmetry graph representation for locomotion policy learning We represent an articulated robot as a graph and complete this graph with morphological symmetry actions on observations, privileged observations, and actions, including the permutations and sign changes induced by sagittal reflection Built on this representation, MS-PPO uses a morphological-symmetry-equivariant graph actor and a morphologicalsymmetry-invariant graph critic The goal is not to define a new RL optimizer, but to define a physically aligned function space in which policy and value functions respect both the robot’s kinematic structure and its morphological symmetry
How the Training Signal is Built
The representation involves augmenting the topological graph with morphological symmetry actions on observations, privileged observations, and actions The actor identity is policy equivariance, meaning mirrored observations produce mirrored actions The critic identity is value invariance, meaning morphologically symmetric states have the same value This is achieved by defining the actor as policy equivariance, where mirrored observations produce mirrored actions <ref:2512.00727#pg2>, and the critic as value invariance, meaning morphologically symmetric states have the same value <ref:2512. p. 300727#pg5> The encoder h(·) maps physical robot signals into graph features that transform only by graph reindexing under G, as in Eq. p. 3 (3)
Where the Gains Come From
Experiments show improved symmetry generalization, robustness, sample efficiency, and model efficiency over topology- and symmetry-aware baselines MS-PPO consistently outperforms strong morphology- and symmetry-aware baselines, achieving the best joint-failure tolerance, the lowest sim-to-real velocity MSE on G1 with only 34.6% of standard PPO’s parameters <ref:2512. p. 100727#pg7> The results show that jointly encoding kinematic topology and morphological symmetry is an effective representation design for locomotion policy learning on articulated robots MS-PPO uses only 34. p. 16% of the parameters of PPO and PPO-Aug and 69. p. 53% of the parameters of PPO-EMLP, suggesting that jointly encoding kinematic connectivity and morphological symmetry improves model efficiency without sacrificing task-level performance <ref:2512. p. 500727#pg7>
Evaluation Across Embodiments
We evaluate MS-PPO on the Unitree Go2 quadruped and G1 humanoid robots MS-PPO achieves the best velocity tracking among all deployable methods, while PPO and MI-PPO fail to produce stable hardware locomotion <ref:2512.00727#pg8> These results support our central claim that effective morphology-aware policy learning requires encoding symmetry structure consistently within the robot’s topological graph representation <ref:2512. p. 600727#pg8> MS-PPO improves symmetry generalization, robustness to asymmetric joint failures, sample efficiency, model efficiency, and zeroshot sim-to-real deployment relative to topology-aware and symmetry-aware baselines
Limitations
Our current validation is limited to bilateral C2 symmetry on one quadruped and one humanoid <ref:2512.00727#pg9> We have not yet tested richer or approximate symmetry structures, such as rotational Cn robots, reflective K4 quadrupeds, radially arranged multi-arm systems, or symmetric manipulators <ref:2512. p. 700727#pg10> Another limitation is that the current experiments instantiate the representation separately on each robot, rather than demonstrating policy transfer across embodiments with different node features, action spaces, and symmetry orbits <ref:2512. p. 700727#pg10> Future work will address these limitations by automating symmetry graph construction and symmetry action deriving <ref:2512. p. 7
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Morphological Symmetry Graph Representation for Locomotion Policy Learning: The system can learn policies that are
symmetry-equivariant
by construction, meaningmirrored observations produce mirrored actions,
which ensures consistent coordination across symmetric body parts. -
Symmetry-Equivariant Graph Actor and Invariant Graph Critic: This architecture enforces the desired policy and value constraints
by construction rather than through regularization or data augmentation,
leading to improved symmetry generalization and robustness over baselines like PPO-EMLP. -
Joint Encoding of Kinematic Topology and Morphological Symmetry: The system achieves better performance by
jointly encoding kinematic connectivity and morphological symmetry,
which is shown to improvemodel efficiency without sacrificing task-level performance
across quadrupedal and humanoid robots. -
Robustness to Asymmetric Joint Failures: By leveraging the graph structure, the system can maintain stable paths even when a joint receives zero torque, as MS-PPO
keeps the failed actuator local to its joint node and lets the topological graph propagate its effect to mechanically related joints.
-
Zero-Shot Sim-to-Real Transfer: The learned policies demonstrate
zero-shot sim-to-real transfer
capability, achieving stable hardware locomotion on a Unitree G1 humanoid with only 34.6% of standard PPO's parameters, showing the representation is effective across embodiments. -
Out-of-Distribution Command Generalization: The system exhibits superior performance in out-of-distribution commands, such as in the Walk-to-One-Side task, where MS-PPO achieves
51.8%, 46.6%, and 61.4% lower RMSE
compared to other methods under these conditions.
Sources
- TWIST: Teleoperated Whole-Body Imitation System
- Towards Dynamic Quadrupedal Gaits: A Symmetry-Guided RL Hierarchy Enables Free Gait Transitions at Varying Speeds
- Coordinated Humanoid Robot Locomotion with Symmetry Equivariant Reinforcement Learning Policy
- Morphologically Symmetric Reinforcement Learning for Ambidextrous Bimanual Manipulation
- Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges
- Proximal Policy Optimization Algorithms
- Symmetry-Guided Memory Augmentation for Efficient Locomotion Learning
- Isaac Lab: A GPU-Accelerated Simulation Framework for Multi-Modal Robot Learning
Related papers
- Rethinking Actions: Continuity-Aware Priors for World-Model Planning
- MOSAIC: Modular Obstacle and Spatial-Aware Framework for Manipulation In Cluttered Scenes
- SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning
- DexSeed: Expanding Human Hand-Object Interactions into Robot Demonstrations for Dexterous Policy Learning
- LaST0.5: A Generalist World Action Model with Interaction-Centric Latent Reasoning
- CIALLO: Learning Shared Intent Representations for Cross-Embodiment Manipulation