Symmetry Equivariant Policy (SE-Policy)

Updated 7 July 2026

SE-Policy is a design principle where policy outputs transform predictably under symmetry group actions, ensuring spatial generalization and robust control.
It employs strategies like invariant encoders, canonicalization, and local-frame parameterization to align observation-action maps with geometric symmetries.
Empirical studies show that SE-Policies improve data efficiency, coordination, and performance across manipulation, locomotion, and path-planning tasks.

A Symmetry Equivariant Policy (SE-Policy) is a policy whose outputs transform predictably under symmetries of the task domain rather than being relearned independently for each transformed configuration. In current robotics and reinforcement-learning literature, the relevant symmetry may be rigid-body motion in three dimensions, planar rotation, or bilateral reflection, depending on the embodiment and task. For manipulation, the dominant case is often $SE(3)$ ; for tabletop control, $SO(2)$ or $SE(2)$ is common; for bimanual and legged systems, $\mathbb{Z}_2$ reflection captures left–right structure. Across these settings, SE-Policies are used to improve spatial generalization, data efficiency, robustness to frame changes, and coordination, and they appear in imitation learning, actor–critic RL, diffusion and flow-based generative control, and symmetry-aware world models (Funk et al., 2024, Zhang et al., 9 Mar 2026, Wei et al., 30 Nov 2025).

1. Formal definition and mathematical structure

The defining condition of an SE-Policy is equivariance. If a group element $g$ acts on observations by $g \cdot x$ and on actions through a representation $\rho(g)$ , a policy $f$ is equivariant when

$f(g \cdot x) = \rho(g) f(x).$

The corresponding invariant condition is $f(g \cdot x) = f(x)$ . In robotic manipulation, the group is frequently

$SO(2)$ 0

with composition

$SO(2)$ 1

acting on points by $SO(2)$ 2 and on poses by left multiplication $SO(2)$ 3 (Funk et al., 2024).

The same formalism appears in simpler or morphology-specific symmetry groups. In bimanual manipulation, EquiBim models bilateral symmetry as a two-element group $SO(2)$ 4 with identity and left–right reflection; in map-based path planning, the exploited symmetry is the cyclic rotation group $SO(2)$ 5; in spatial-action Q-learning, the relevant subgroup is a discretized form of $SO(2)$ 6; and in humanoid or quadruped locomotion, bilateral reflection is realized by permutations and sign flips over state and action coordinates (Zhang et al., 9 Mar 2026, Theile et al., 2024, Wang et al., 2021, Nie et al., 2 Aug 2025).

For reinforcement learning, the formal basis is a symmetric MDP. When reward and transition satisfy symmetry-compatible conditions such as

$SO(2)$ 7

the optimal value is invariant and the optimal policy is equivariant. This argument underlies MDP homomorphic networks, equivariant Q-learning, symmetry-aware actor–critic locomotion, and symmetry-constrained world-model control (Pol et al., 2020, Wang et al., 2021, Wei et al., 30 Nov 2025, Lan et al., 18 Jun 2026).

A central implication is that global transformations of the scene or embodiment should not require new demonstrations or distinct control modes. In the manipulation setting, this is often expressed as “spatial generalization”; in locomotion, as “symmetry generalization” or mirrored coordination; in path planning, as transfer across rotated maps. The mathematical object is the same, but the representation $SO(2)$ 8 differs with the action space.

2. Representation strategies: invariance, canonicalization, and local frames

Recent SE-Policy designs differ primarily in how they obtain symmetry-aware internal representations. One line uses invariant encoders over relative geometry. ActionFlow introduces an $SO(2)$ 9 Invariant Transformer in which observations and actions are tokenized as poses and semantic features, and attention is built from invariant relative-pose quantities, distances, inner products, and Invariant Point Attention. This yields a latent representation invariant to global frame changes while retaining strong locality biases (Funk et al., 2024).

A second line uses canonicalization. Canonical Policy learns a canonical transform $SE(2)$ 0 that maps a 3D point cloud to a canonical frame and then predicts actions in that frame, with the overall policy recovered by transporting canonical actions back into the world frame. EquiForm adopts the same canonicalization perspective but adds a non-learned geometric denoising module and a contrastive equivariant alignment loss to stabilize the transform estimate and feature space under noise, occlusion, and partial visibility (Zhang et al., 24 May 2025, Zhang et al., 24 Jan 2026).

A third strategy relies on local observations and local action coordinates. “A Practical Guide for Incorporating Symmetry in Diffusion Policy” proves that combining eye-in-hand perception with relative or delta action parameterization yields inherent $SE(2)$ 1-invariance at the local action level, with equivariance recovered when absolute actions are reconstructed through the current end-effector pose. EquiContact uses an analogous principle in contact-rich control: its low-level policy operates on geometrically consistent error vectors, force–torque readings, and wrist RGB features expressed in the end-effector frame, making the local policy left-invariant while the composed spatial action becomes equivariant (Wang et al., 19 May 2025, Seo et al., 15 Jul 2025).

A fourth strategy constructs equivariance directly from monocular RGB. “3D Equivariant Visuomotor Policy Learning via Spherical Projection” lifts 2D eye-in-hand image features onto $SE(2)$ 2, aligns them with the world frame through an end-effector-orientation correction, and processes them with spherical and $SE(2)$ 3-equivariant layers. This produces a closed-loop visuomotor policy that is globally $SE(2)$ 4-equivariant without explicit point-cloud reconstruction (Hu et al., 22 May 2025).

These designs clarify an important point: policy-level equivariance does not require a single universal recipe. Some methods enforce equivariance through globally equivariant backbones; others through invariant canonicalization, local-frame parameterization, or symmetry-aware feature aggregation. This suggests that SE-Policy is better understood as a property of the end-to-end observation-to-action map than as a commitment to one architectural family.

3. Generative SE-Policies: diffusion, flow matching, and compliant decoding

A large recent subliterature realizes SE-Policies through generative models over action sequences. The dominant pattern is to combine a symmetry-aware encoder with a decoder whose denoising or flow updates respect the action-space group structure.

Equivariant Diffusion Policy analyzes $SE(2)$ 5 symmetry for full 6-DoF control and constructs a denoiser satisfying

$SE(2)$ 6

using equivariant observation and action encoders, a shared temporal U-Net over a discrete rotation subgroup, and an equivariant decoder into action irreducible representations. The forward noising kernel is isotropic, so the reverse process remains equivariant in distribution (Wang et al., 2024).

ET-SEED moves from single actions to trajectory-level $SE(2)$ 7 diffusion on the manifold itself. Its key theoretical move is to relax the usual demand that every reverse-diffusion transition be equivariant: the first $SE(2)$ 8 denoising steps are condition-invariant, and only the final step is output-equivariant, yet the final trajectory distribution still satisfies the desired $SE(2)$ 9 transformation law. This reduces training difficulty while preserving end-to-end symmetry (Tie et al., 2024).

ActionFlow replaces diffusion with Conditional Flow Matching. Actions are sequences of $\mathbb{Z}_2$ 0 end-effector poses, and the decoder updates translation and rotation in each action’s local frame:

$\mathbb{Z}_2$ 1

Because translation is applied through the current local rotation and rotations are integrated on the Lie group, the decoder implements the natural $\mathbb{Z}_2$ 2 action on outputs when driven by an invariant latent representation (Funk et al., 2024).

ReSeFlow further rectifies this direction by replacing iterative denoising with rectified flows in $\mathbb{Z}_2$ 3 and a right-invariant ODE,

$\mathbb{Z}_2$ 4

trained toward geodesic displacements. The paper reports that one inference step can outperform a 100-step baseline on its simulated benchmarks, linking equivariance to inference efficiency as well as data efficiency (Wang et al., 20 Sep 2025).

EquiContact broadens the scope from pose generation to vision-to-force control. Its high-level planner estimates an object-centric reference frame, while its low-level compliant policy outputs relative poses and admittance gains that feed a geometric admittance controller. The resulting wrench transformation follows the co-adjoint action, and the full system is formulated as an end-to-end $\mathbb{Z}_2$ 5-equivariant mapping from perception to compliant force control (Seo et al., 15 Jul 2025).

Taken together, these works show that “SE-Policy” in generative control is not merely a property of the score network. It depends on the symmetry of the latent encoder, the trajectory parameterization, the Lie-group update rule, and, in contact-rich settings, the control law that interprets generated poses or wrenches.

4. Reinforcement-learning formulations: equivariant actors, invariant critics, and symmetric world models

In RL, SE-Policies appear in several distinct forms. One is value-based control in spatial action spaces. “Equivariant $\mathbb{Z}_2$ 6 Learning in Spatial Action Spaces” shows that under reward and transition symmetry, the optimal $\mathbb{Z}_2$ 7 function satisfies

$\mathbb{Z}_2$ 8

which implies an equivariant greedy policy $\mathbb{Z}_2$ 9. The paper implements this with steerable convolutional Q-networks over discretized $g$ 0 action maps and reports better sample efficiency on manipulation benchmarks (Wang et al., 2021).

A second form replaces special layers with exact symmetrization. “Equivariant Ensembles and Regularization for Reinforcement Learning in Map-based Path Planning” constructs an exactly equivariant policy by averaging over transformed inputs and undoing action permutations,

$g$ 1

and an invariant value by averaging over transformed states. This preserves exact symmetry at the ensemble level while using standard neural components (Theile et al., 2024).

A third form builds symmetry directly into actor–critic architectures for embodied control. “Leveraging Symmetry in RL-based Legged Locomotion Control” studies a strictly equivariant actor and invariant critic for sagittal reflection in quadrupeds, contrasted with data-augmentation approximations. MS-PPO extends this to morphology-informed graph neural networks, proving actor equivariance and critic invariance through composition of a symmetry-equivariant encoder, an equivariant graph backbone, and an equivariant or invariant readout. “Coordinated Humanoid Robot Locomotion with Symmetry Equivariant Reinforcement Learning Policy” applies the same principle to a 27-DoF humanoid with exact bilateral transforms on IMU signals, commands, joint states, action histories, phase variables, and critic-only terrain maps (Su et al., 2024, Wei et al., 30 Nov 2025, Nie et al., 2 Aug 2025).

A fourth form embeds symmetry into latent dynamics. SWAP trains a symmetric equivariant world model and a high-frequency equivariant actor with an invariant critic. The latent parity operator swaps left–right latent pairs, and the recurrent model, prior, posterior, and decoder all satisfy symmetry-compatible transformation equations. This makes the latent state itself an equivariant object consumed by the policy (Lan et al., 18 Jun 2026).

Finally, symmetry can govern post hoc RL fine-tuning of generative policies. “Symmetry-Aware Steering of Equivariant Diffusion Policies” proves that an equivariant diffusion generator induces a group-invariant latent-noise MDP. Steering is then performed by an RL actor over latent noise and an invariant critic over latent state–noise pairs. The paper contrasts strict equivariant steering, approximate equivariant steering, and non-equivariant steering, arguing that the choice should depend on how strongly the downstream task preserves the assumed symmetry (Park et al., 12 Dec 2025).

An enduring conceptual pattern is that RL work nearly always pairs an equivariant policy with an invariant value estimator. This is not incidental: action selection must transform with symmetry, whereas the expected return assigned to symmetric states should coincide.

5. Empirical landscape across domains

The empirical record for SE-Policies is broad rather than uniform. Reported gains differ with the task symmetry, observation modality, action parameterization, and the degree to which real dynamics violate the assumed group action. The table summarizes representative results drawn directly from papers.

Setting	Symmetry and formulation	Reported result
ActionFlow (Funk et al., 2024)	$g$ 2-invariant transformer + Flow Matching	Real-time action generation of $g$ 3– $g$ 4 s per sequence on RTX 3090; lightbulb mounting 8/10; mug hanging test 7/10 with point cloud token and 8/10 with pose conditioning
Equivariant Diffusion Policy (Wang et al., 2024)	$g$ 5-equivariant diffusion for 6-DoF control	Average success 63.9% vs 42.0% with 100 demos; 72.6% vs 57.8% with 200 demos; 77.9% vs 71.4% with 1000 demos
ET-SEED (Tie et al., 2024)	Trajectory-level $g$ 6 diffusion with relaxed kernel condition	Open Bottle Cap, new poses, 25 demos: 74% ± 6.52; Open Door, new poses, 25 demos: 66% ± 2.74
EquiBim (Zhang et al., 9 Mar 2026)	$g$ 7 bilateral equivariance regularization for dual-arm IL	Image + Joint on RoboTwin: 34.1 → 43.6; Banana shifted distribution on real robot: ACT 0/10 → ACT+EquiBim 5/10
Canonical Policy (Zhang et al., 24 May 2025)	Learned canonical 3D representation for equivariant IL	Average improvement of 18.0% in simulation and 37.6% in real-world experiments
EquiForm (Zhang et al., 24 Jan 2026)	Canonicalization + geometric denoising + contrastive equivariant alignment	Average improvement of 17.2% in simulation and 28.1% in real-world experiments
MS-PPO (Wei et al., 30 Nov 2025)	Morphological-symmetry-equivariant actor–critic for locomotion	Trot RMSE-O reductions of 51.8%, 46.6%, and 61.4% versus PPO-MLP, PPO-EMLP*, and MI-PPO
SWAP (Lan et al., 18 Jun 2026)	Symmetric equivariant world model + actor–critic	Real robot gap leap of 2.13 m and platform climb of 1.63 m
Equivariant Push-Grasp Network (Hu et al., 3 Apr 2025)	$g$ 8-equivariant push–grasp score maps	Grasp success improvements of 49% in simulation and 35% in real-world scenarios compared to strong baselines

These results cover at least five distinct use cases. First, spatially generalizable manipulation is the most mature domain: ActionFlow, ET-SEED, Canonical Policy, EquiForm, Equivariant Diffusion Policy, and the Equivariant Push-Grasp Network all report gains tied explicitly to geometric variation in object pose, scene layout, or camera viewpoint. Second, contact-rich control now includes explicitly equivariant vision-to-force pipelines, as shown by EquiContact’s near-perfect peg-in-hole generalization with a small demonstration set. Third, bilateral symmetry is now exploited directly in bimanual imitation learning through EquiBim’s model-agnostic training-time regularizer. Fourth, locomotion work consistently links symmetry constraints to more coordinated and periodic gaits, improved robustness, and zero-shot sim-to-real transfer. Fifth, path-planning and map-based RL demonstrate that even discrete rotational symmetry can materially improve convergence and performance (Seo et al., 15 Jul 2025, Zhang et al., 9 Mar 2026, Wei et al., 30 Nov 2025, Theile et al., 2024).

The empirical record also shows that symmetry benefits are largest in low-data or distribution-shifted regimes. Equivariant Diffusion Policy outperforms its strongest baseline by 21.9% on average with 100 demonstrations; EquiBim’s largest simulation gains appear in Image + Joint configurations, where geometry priors are weakest; Canonical Policy and EquiForm emphasize out-of-distribution robustness to pose and viewpoint changes; and the steering study finds the biggest gains when equivariant diffusion policies are fine-tuned from extremely limited demonstrations (Wang et al., 2024, Zhang et al., 9 Mar 2026, Zhang et al., 24 May 2025, Zhang et al., 24 Jan 2026, Park et al., 12 Dec 2025).

6. Limitations, misconceptions, and open directions

The main limitation of SE-Policies is not architectural but epistemic: symmetry assumptions may be only approximately true. Several papers explicitly warn that depth noise, calibration errors, occlusion, appearance changes, friction anisotropy, non-rigid deformation, local task asymmetries, or role-asymmetric coordination can break the assumed group action and reduce the benefit of strict equivariance (Funk et al., 2024, Zhang et al., 9 Mar 2026, Zhang et al., 24 Jan 2026, Seo et al., 15 Jul 2025).

This leads to an active design question: when should symmetry be enforced strictly, softly, or only approximately? Strict approaches include exact equivariant backbones, exact actor equivariance with invariant critics, and exact ensemble symmetrization. Softer approaches include prediction-level regularizers, symmetry-aware data augmentation, and approximate equivariant steering with residual non-equivariant branches. Recent comparisons indicate that strict equivariance tends to dominate when task symmetry is strong, but approximate methods can be more robust when joint limits, occlusions, workspace constraints, or asymmetric demonstrations break the symmetry in practice (Su et al., 2024, Park et al., 12 Dec 2025, Zhang et al., 9 Mar 2026).

A common misconception is that equivariance is synonymous with specialized steerable networks. The literature shows several alternatives: exact symmetrization by ensembles, invariant local action representations, frame averaging over pretrained encoders, canonicalization into a learned pose-normalized frame, and symmetry-aware training-time penalties can all yield policy-level SE behavior without making every block explicitly equivariant (Theile et al., 2024, Wang et al., 19 May 2025, Zhang et al., 24 May 2025).

Another misconception is that symmetry necessarily improves every metric. The evidence is narrower. Gains are strongest when the symmetry matches the physical task and the observation/action pipeline preserves it. Where symmetry is weak, partially broken, or confounded by sensing, enforced equivariance can underfit useful asymmetries. The steering paper documents this directly; legged-locomotion studies note real-world morphology and environment mismatches; and canonicalization papers emphasize that severe occlusion can violate the assumption that new observations differ by rigid transforms alone (Park et al., 12 Dec 2025, Su et al., 2024, Zhang et al., 24 May 2025).

Open directions are already visible in the papers themselves. These include extending from $g$ 9 or bilateral reflection to full $g \cdot x$ 0 in more settings, learning object-specific isotropy groups, combining morphology and task symmetries, using adaptive step sizing and broader real-world validation for rectified-flow policies, handling multi-object and partially observed scenes, and integrating symmetry-aware policies with world models and contact-adaptive control (Wang et al., 20 Sep 2025, Zhang et al., 24 May 2025, Zhang et al., 24 Jan 2026, Lan et al., 18 Jun 2026).

In that sense, SE-Policy has evolved from a narrow architectural constraint into a general design principle: identify the symmetry group that the task, embodiment, and sensing pipeline genuinely support; encode it at the right level of abstraction; and let the policy, critic, or generative decoder inherit that structure. The current literature shows that when this alignment is achieved, the result is not only cleaner group-theoretic behavior, but also measurable gains in generalization, data efficiency, coordination, and real-time control.