Pith. sign in

REVIEW 6 major objections 6 minor 1 cited by

SE(3)-Equivariant Diffusion Policy in Spherical Fourier Space

T0 review · 6 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read This paper introduces a diffusion policy that is end-to-end equivariant under 3D rotations and translations by carrying states, actions, and the denoising process in spherical Fourier space, so that rotating or shifting the scene…

desk verdict Strong empirical paper with a load-bearing equivariance claim that is under-proved and under-tested; fixable. read the letter →

arxiv 2507.01723 v1 pith:SZFGHZ2W submitted 2025-07-02 cs.RO

classification cs.RO
keywords SE(3)equivariancesphericalFourierfeaturesdiffusionpolicyrobotmanipulationimitationlearningclosed-loopbi-manualharmonics
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper introduces Spherical Diffusion Policy (SDP), a closed-loop robot manipulation policy that is equivariant to the full SE(3) group of 3D rotations and translations. Its purpose is to make policies trained from a small number of human demonstrations generalize to new 3D arrangements of objects without collecting more data. The method embeds states, actions, and the denoising process in spherical Fourier space, where a rotation acts linearly through Wigner D-matrices, and conditions the denoiser through spherical FiLM layers. Translation invariance comes from a relative action formulation that centers observation and action on the gripper. The authors prove that if the optimal policy is SE(3)-equivariant, the entire denoising function is equivariant by composition, and they report higher success rates than the baselines across twenty simulation tasks and five physical robot tasks, including bi-manual manipulation.

What carries the argument

Spherical Fourier space: states, actions, and noise are represented as spherical harmonic coefficients up to degree $L$, under which a rotation $g$ acts by the orthogonal Wigner D-matrix $D^l(g)$. The paper constructs a spherical denoising temporal U-net (SDTU) whose mixing-channel 1D convolutions act independently on each degree $l$, preserving equivariance over time, and spherical FiLM (SFiLM) layers that modulate each degree's features with a rotated scaling and offset. These pieces make the denoising function equivariant by composition; translation invariance comes from a relative action formulation that defines positions with respect to the gripper.

What would settle it

Run SDP and a non-equivariant baseline on a gravity-sensitive task, such as pouring from a cup or placing an object with a required upright orientation, with the object at many 3D poses; if SDP's success rate falls with orientation as steeply as, or more steeply than, the baseline's, then the SE(3) equivariance assumption is not the source of the observed generalization.

Watch

Extended reading notes

Core claim

The central claim is that SDP is end-to-end SE(3)-equivariant: rotating the input scene rotates both the internal spherical Fourier features and the final denoised action trajectory, while translating the scene leaves the trajectory unchanged. The authors build this from three equivariant components — a spherical encoder, a spherical denoising temporal U-net, and spherical FiLM conditioning — and the equivariance of the composition is proven from Schur's lemma and the orthogonality of the Wigner D-matrices. Because the symmetry is hard-wired rather than learned, the policy can transfer to transformed scenes without additional training data. In their experiments, SDP reaches an average physical-task success rate of 78 percent against 17 and 7 percent for the two strongest baselines, and it outperforms all baselines on most simulation tasks with SE(3) and SE(2) initialization.

Load-bearing premise

The load-bearing premise is that the optimal manipulation policy is exactly SE(3)-equivariant, which fails whenever gravity, contact forces, occlusion, or unstable objects make the orientation of the scene matter for the task.

Editorial extensions

If this is right

  • Closed-loop manipulation policies trained on a handful of demonstrations should generalize to novel object poses and table orientations without fine-tuning.
  • Continuous SO(3) equivariance avoids the discretization error that limits discrete-symmetry baselines, so the policy handles arbitrary rotation angles rather than only multiples of a fixed step.
  • Because positions are defined relative to the gripper, the policy is insensitive to translations of the scene or robot base, easing deployment across robot stations.
  • The compact degree-2 spherical representation keeps inference around 0.44 seconds, compatible with closed-loop control, while preserving the equivariance guarantee.
  • The same relative-action and spherical-representation recipes extend to bi-manual robots by canonicalizing to each gripper, as demonstrated on three physical bi-manual tasks.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural stress test is to weaken the symmetry to per-task subgroups (for example, SE(2) on a flat table) and compare with full SE(3): tasks dominated by gravity should favor the weaker symmetry, and this comparison would clarify when hard-wired equivariance helps or hurts.
  • The spherical Fourier machinery is not specific to diffusion; the same encoder and SFiLM conditioning could be plugged into deterministic behavior cloning or reinforcement learning, which would test whether the equivariance or the diffusion modeling drives the sample-efficiency gain.
  • The tilting-table results suggest an implicit limitation: occlusion can hide part of the scene rotation, so a policy that combines equivariance with explicit reasoning about unobserved geometry might outperform SDP on highly tilted scenes.
  • Since translation invariance is enforced by canonicalizing to the gripper, tasks where the absolute position in the workspace matters, such as placing an object onto a fixed target or avoiding a wall, are a direct test of the assumption's boundaries.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

6 major / 6 minor

Summary. The paper proposes Spherical Diffusion Policy (SDP), a closed-loop imitation learning policy that represents states, actions, and denoising latents in spherical Fourier space and claims end-to-end SE(3) equivariance. Equivariance is built from an EquiformerV2 point-cloud encoder, a spherical denoising temporal U-net (SDTU) with m-independent spherical Fourier convolutions, and newly proposed spherical FiLM (SFiLM) conditioning; translation invariance is obtained through a relative action formulation. The paper reports success rates on 20 simulated MimicGen tasks and 5 physical robot tasks, including bimanual tasks, and compares against EquiDiff, DiffPo, DP3, EquiBot, and other baselines.

Significance. If the central equivariance claim holds, the paper makes a substantial contribution: it offers a compact spherical-Fourier alternative to heavy SO(3) irreducible representations, with a modular design whose equivariance composes across encoder, denoiser, and conditioning layers, and it demonstrates strong gains in data efficiency and real-robot generalization, including bimanual manipulation. The public code release and the breadth of the simulation and physical evaluation are assets. However, the proof of the key convolution equivariance is misstated, the implementation's exact equivariance is never numerically verified, and some empirical claims are contradicted by the paper's own tables, so the significance is conditional on correction of these issues.

major comments (6)
  1. [§4.3, Eq. (4), Appendix A.1] Proposition 4.1 as stated in Eq. (4) is not the equivariance condition for the mixing convolution: it asserts that multiplying both sides of Eq. (3) by the same Wigner matrix element D^l_mn(g) yields equality, and the proof in A.1 simply factors this scalar out of the sums. The true equivariance requirement is that the rotated output equals the convolution evaluated on the rotated input, which requires summing over m and commuting the Wigner D-matrix with the channel-time weights. The operation in Eq. (3) can indeed be made equivariant because the weights w^{i,o}_{l,j-t} are independent of m, but the paper never states this cleanly, and the displayed proof does not establish it. Since Section 4.1's 'equivariant by composition' argument and the abstract's central claim rest on this proposition, the statement and proof need to be corrected.
  2. [§5 (all experiments)] No experiment verifies the defining equivariance property. For a fixed noise seed, one could compare epsilon_theta(gS, gA, k) with g epsilon_theta(S, A, k), or compare final action trajectories A(gS) with g A(S). Such a test would detect implementation-level non-equivariance introduced by normalization layers in EquiformerV2, spherical activation, point-cloud downsampling, or the division by ||h_l|| in SFiLM. Without it, the claim of exact end-to-end SE(3) equivariance is supported only by the (currently flawed) proof, not by the implementation, and the reported gains against non-equivariant baselines cannot rule out an approximate implementation.
  3. [§5.4, Table 5] The claim that 'beyond degree 2, performance saturates' is contradicted by Table 5: degree 3 gives average success 56% versus 49% for degree 2, and on Square 15° the gap is 66% versus 38%. If the intended message is that degree 2 offers a better accuracy/cost trade-off for real-time control, that should be stated explicitly; as written, the degree-saturation claim is unsupported by the reported numbers.
  4. [§5.2, §5.3, §5.4 (Tables 3, 4, 5)] The physical experiments, the ablation study, and the degree study are each reported from a single seed without variance. Given the large differences in Table 4 (e.g., 49% vs. 35% for Discrete SDP, and 49% vs. 26% for SDP Abs.), it is not possible to assess whether these differences are statistically significant. At least for the simulation ablations, which are comparatively cheap, multiple seeds should be reported; for the physical results, variance across rollouts or an explicit statement of the one-seed limitation is needed before the 'large improvement' claim can be considered robust.
  5. [§5.1, Table 1] The simulation protocol trains and evaluates within the same tilt ranges, so the reported gains demonstrate within-distribution robustness, not generalization to rotations outside the training support. Since the abstract claims 'robust generalization across transformed 3D scenes,' a held-out evaluation (for example, training on [0, 15°] and evaluating on [15°, 30°]) would directly test whether the equivariance constraint transfers beyond the training distribution. Without such an evaluation, the generalization claim is weaker than stated.
  6. [§5.1, §5.4, Tables 1-3, Table 6] The closest SE(3)-equivariant diffusion policy baseline, ET-SEED (Tie et al., 2024), is absent from all performance comparisons; the paper only compares inference time and batch size in Table 6. Because the contribution is explicitly positioned against ET-SEED's heavy SO(3) irreducible representations and two-stage inference, the empirical claim of 'large performance improvement over strong baselines' is incomplete without a direct comparison. The stated reason (code unavailable before submission) is understandable, but the claim should be tempered, or a later comparison added.
minor comments (6)
  1. [§4.4, Eq. (5)] The SFiLM operation divides by ||h_l|| and is undefined when h_l = 0; the equivariance statement should either exclude this case or specify a convention for the zero vector.
  2. [§5.3] The phrase 'sufferers from discretization error' should read 'suffers from discretization error.'
  3. [§5.3, ablation 3] The DP3-canonical ablation is attributed to 'Ke et al., 2024,' but the DP3 baseline elsewhere in the paper is attributed to Ze et al., 2024; please correct the citation.
  4. [§5.2, Table 3] The entry '35/ 90*' for Push Eraser is visually ambiguous; clarify that 35% corresponds to 1024 points and 90% to 2048 points, and explain the formatting of the asterisk in a footnote.
  5. [§5.1] The sentence 'As the degress of SE(3) initialization increases' contains a typo: 'degress' should be 'degrees.'
  6. [Appendix A.2] The justification 'the line 16 is based on Schur's lemma' is imprecise; line 16 follows from orthogonality of the Wigner D-matrix and from linearity, not from Schur's lemma, which is unnecessary for this scalar projection argument.

Circularity Check

1 steps flagged · score 4.0 of 10

The Proposition 4.1 equivariance proof is tautological (Eq. 4 re-states Eq. 3 with a common Wigner factor), but the architecture itself is genuinely equivariant and the empirical results remain independent; overall circularity is limited.

  1. self definitional [Section 4.3, Proposition 4.1 (Eq. 4), proof in Appendix A.1]
    "Proposition 4.1. The mixing channel temporal convolution in Equation. 3 is SO(3) equivariant: D^l_mn(g)h^o_{l,m,t} = Σ_{j∈T} Σ_{i∈in} D^l_mn(g)h^i_{l,m,j}w^{i,o}_{l,j-t}. ... = D^l_mn(g) h^o_{l,m,t}, the line 8 is because of Schur's lemma, which states that any linear operation of SO(3) irreps acts as on each irreducible subspace is equivariant."

    The claimed equivariance condition (4) is obtained from the convolution definition (3) by multiplying both sides by the same scalar D^l_mn(g); it never states how a rotated feature (a sum over orders) maps to the output. The proof in A.1 factors out D^l_mn(g), then invokes Schur's lemma to assert the very equivariance being proved. The derivation therefore reduces to a tautology: Eq. 4 is Eq. 3 with a common Wigner factor, not a transformation law. The operation in Eq. 3 is in fact equivariant because the weights w are independent of the order m, but the paper does not show this. Since Section 4.1 uses this proposition to conclude 'the denoising function is equivariant by composition,' the theoretical validation is unsupported as written.

full rationale

Aside from the tautological Proposition 4.1 proof, the derivation chain is not circular. The equivariant-diffusion link is cited to Ryu et al. (2024) and Wang et al. (2024b, with overlapping author Walters), but Appendix C.1 re-proves it in a self-contained way, so the self-citation is not load-bearing. Translation invariance is explicitly enforced by canonicalizing state and action via the relative action formulation, and the paper presents this as a construction rather than as a predicted outcome. The empirical generalization results compare against strong baselines on held-out poses and are independent of any fitted parameter. The main circular step is confined to the proof of Proposition 4.1, where the equivariance equation is Eq. (3) multiplied by D^l_mn(g) and the proof invokes Schur's lemma to restate the claim. Because the convolution operation itself is actually equivariant (weights are m-independent), the architecture's claim is likely salvageable, but the paper's own derivation of the SDTU component does not establish it. No numerical equivariance test (e.g., fixed-seed A(gS)=gA(S)) is reported, which would have independently verified the property; its absence is a verification gap rather than a circular fit. Overall, one proof step is circular/tautological, while the central method and evaluations retain independent content.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a domain assumption about task equivariance, a dependence on the cited EquiformerV2 encoder, and standard mathematical facts. The only hand-chosen hyperparameter with a direct performance effect is the spherical harmonic degree, which the paper's own data show is not set at the accuracy optimum.

free parameters (2)
  • Maximum spherical harmonic degree L = 2
    Chosen by hand as a trade-off between performance and inference cost; Table 5 shows degree 3 achieves 56% average vs 49% for degree 2, so the choice is not optimal on accuracy.
  • Point cloud resolution = 1024 points (2048 for one Push Eraser variant)
    Resolution is a model choice; the paper notes higher resolution improves Push Eraser by over 50%, so performance is sensitive to this hyperparameter.
assumptions (4)
  • domain assumption The optimal manipulation policy is SE(3) equivariant.
    Section 4.1 states this assumption directly. It is load-bearing because the entire architecture is constrained to it, yet gravity and occlusion break it in tilted scenes (Section 5.1).
  • domain assumption EquiformerV2 provides an SE(3)-equivariant point cloud encoder.
    Section 4.2 uses EquiformerV2 for the spherical encoder, relying on prior work for its equivariance properties.
  • domain assumption The denoising function is equivariant if the policy is equivariant and the noise distribution is invariant.
    Taken from Ryu et al. (2024) and Wang et al. (2024b), summarized in Appendix C.1. The paper treats this as a foundation, not as something derived in this work.
  • standard math Wigner D-matrices are orthogonal and Schur's lemma applies to the spherical Fourier coefficients.
    Used in the proofs of Propositions 4.1 and 4.2 in Appendix A.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SE(3)-Equivariant Diffusion Policy in Spherical Fourier Space." pith.science (2026). https://pith.science/paper/SZFGHZ2W

@misc{pith2026250701723,
  author       = {Pith},
  title        = {Pith review of: SE(3)-Equivariant Diffusion Policy in Spherical Fourier Space},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SZFGHZ2W}},
  note         = {Machine review of arXiv:2507.01723}
}
read the original abstract

Diffusion Policies are effective at learning closed-loop manipulation policies from human demonstrations but generalize poorly to novel arrangements of objects in 3D space, hurting real-world performance. To address this issue, we propose Spherical Diffusion Policy (SDP), an SE(3) equivariant diffusion policy that adapts trajectories according to 3D transformations of the scene. Such equivariance is achieved by embedding the states, actions, and the denoising process in spherical Fourier space. Additionally, we employ novel spherical FiLM layers to condition the action denoising process equivariantly on the scene embeddings. Lastly, we propose a spherical denoising temporal U-net that achieves spatiotemporal equivariance with computational efficiency. In the end, SDP is end-to-end SE(3) equivariant, allowing robust generalization across transformed 3D scenes. SDP demonstrates a large performance improvement over strong baselines in 20 simulation tasks and 5 physical robot tasks including single-arm and bi-manual embodiments. Code is available at https://github.com/amazon-science/Spherical_Diffusion_Policy.

Figures

Figures reproduced from arXiv: 2507.01723 by the authors.

Figure 1
Figure 1. SDP enforces that the policy is SO(3) equivariant. Specifically, in the second row, an SO(3) rotation that is applied to the scene leads to an equivalent rotation on the latent spherical Fourier features in the neural networks enc, ϵθ, and on the gen￾erated trajectory (blue dots). Fourier features are visualized as spherical signal. Consider, for example, grasping a dish that is randomly placed in the sink, threadin… view at source ↗
Figure 2
Figure 2. Method overview. During inference, SDP first embeds state St into a spherical scene feature Ct by the encoder enc. Then, SDTU ϵθ estimates the noise ϵ based on the noisy actions A k t , step k, and the scene feature Ct. Later, the noise is subtracted from the noisy actions, generating cleaner actions A k−1 t . This denoising process is performed for K iterations, generating a clean trajectory A 0 t . 4. Method 4.1. … view at source ↗
Figure 3
Figure 3. Spherical denoising temporal U-net (SDTU). Left: The SDTU ϵθ estimates the noise ϵ, based on the noisy actions A k t , denoising step index k, and the encoded scene C. The SDTU has a U-net architecture, with 4 spherical down or up convolution blocks. Right: details of a spherical down or up convolution block. We adopt point clouds as an observation o and treat color information as degree 0 spherical coefficients (sa… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: MimicGen tasks with SE(3) initialization ((a)-(c), showing 1 of 4 tasks) and SE(2) initialization ((d)-(f), showing 3 of 12 tasks) [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Five physical robotic manipulation tasks. Push Eraser we also tested 2048 points). The actions are the 6 DoF gripper poses for single-arm tasks or 12 DoF gripper poses for bi-manual tasks. Training Dataset from Human Demonstrations We use Gello (Wu et al., 2023) to col…
Figure 6
Figure 6. Figure 6: Impact of training dataset size on task success rates: in￾creasing the number of demonstrations from 100 to 316 (a 3× increase) yields an average success rate improvement of 9–12% across four tasks, with initial success rates of 48.5% (SDP), 20.3% (EquiDiff), and 8.8% …

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Static In, Dynamic Out: Counterfactual Action Augmentation for Moving Object Manipulation

    cs.RO 2026-07 conditional novelty 6.0 of 10

    SIDO morphs static demonstrations into counterfactual future-pose samples, training a goal-conditioned policy that, paired with a pose predictor, grasps objects whose motion was unseen during training.

Reference graph

Works this paper leans on

74 extracted references · 49 canonical work pages · cited by 1 Pith paper

  1. [1]

    Spherical fourier neural operators: Learning stable dynamics on the sphere

    Bonev, B., Kurth, T., Hundt, C., Pathak, J., Baust, M., Kashinath, K., and Anandkumar, A. Spherical fourier neural operators: Learning stable dynamics on the sphere. In International conference on machine learning, pp.\ 2806--2823. PMLR, 2023

  2. [2]

    Brehmer, J., Bose, J., De Haan, P., and Cohen, T. S. Edgi: Equivariant diffusion for planning with embodied agents. Advances in Neural Information Processing Systems, 36, 2024

  3. [3]

    Diffusion policy: Visuomotor policy learning via action diffusion

    Chi, C., Feng, S., Du, Y., Xu, Z., Cousineau, E., Burchfiel, B., and Song, S. Diffusion policy: Visuomotor policy learning via action diffusion. In Proceedings of Robotics: Science and Systems (RSS), 2023

  4. [4]

    Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots

    Chi, C., Xu, Z., Pan, C., Cousineau, E., Burchfiel, B., Feng, S., Tedrake, R., and Song, S. Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots. In Proceedings of Robotics: Science and Systems (RSS), 2024

  5. [5]

    S., Geiger, M., K \"o hler, J., and Welling, M

    Cohen, T. S., Geiger, M., K \"o hler, J., and Welling, M. Spherical cnns. In International Conference on Learning Representations, 2018

  6. [6]

    Fin ray® effect inspired soft robotic gripper: From the robosoft grand challenge toward optimization

    Crooks, W., Vukasin, G., O’Sullivan, M., Messner, W., and Rogers, C. Fin ray® effect inspired soft robotic gripper: From the robosoft grand challenge toward optimization. Frontiers in Robotics and AI, 3, 2016. ISSN 2296-9144. doi:10.3389/frobt.2016.00070. URL https://www.frontiersin.org/journals/robotics-and-ai/articles/10.3389/frobt.2016.00070

  7. [7]

    Deng, C., Litany, O., Duan, Y., Poulenard, A., Tagliasacchi, A., and Guibas, L. J. Vector neurons: A general framework for so (3)-equivariant networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 12200--12209, 2021

  8. [8]

    A., Wahid, A., Downs, L., Wong, A., Lee, J., Mordatch, I., and Tompson, J

    Florence, P., Lynch, C., Zeng, A., Ramirez, O. A., Wahid, A., Downs, L., Wong, A., Lee, J., Mordatch, I., and Tompson, J. Implicit behavioral cloning. In Conference on Robot Learning, pp.\ 158--168. PMLR, 2022

Show all 74 references
  1. [9]

    Riemann: Near real-time se (3)-equivariant robot manipulation without point cloud segmentation

    Gao, C., Xue, Z., Deng, S., Liang, T., Yang, S., Shao, L., and Xu, H. Riemann: Near real-time se (3)-equivariant robot manipulation without point cloud segmentation. arXiv preprint arXiv:2403.19460, 2024

  2. [10]

    and Smidt, T

    Geiger, M. and Smidt, T. e3nn: Euclidean neural networks. arXiv preprint arXiv:2207.09453, 2022

  3. [11]

    Deep residual learning for image recognition

    He, K., Zhang, X., Ren, S., and Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 770--778, 2016

  4. [12]

    Denoising diffusion probabilistic models

    Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 0 6840--6851, 2020

  5. [13]

    G., Vignac, C., and Welling, M

    Hoogeboom, E., Satorras, V. G., Vignac, C., and Welling, M. Equivariant diffusion for molecule generation in 3d. In International conference on machine learning, pp.\ 8867--8887. PMLR, 2022

  6. [14]

    Adaptive compliance policy: Learning approximate compliance for diffusion guided control

    Hou, Y., Liu, Z., Chi, C., Cousineau, E., Kuppuswamy, N., Feng, S., Burchfiel, B., and Song, S. Adaptive compliance policy: Learning approximate compliance for diffusion guided control. arXiv preprint arXiv:2410.09309, 2024

  7. [15]

    Orbitgrasp: Se (3)-equivariant grasp learning

    Hu, B., Zhu, X., Wang, D., Dong, Z., Huang, H., Wang, C., Walters, R., and Platt, R. Orbitgrasp: Se (3)-equivariant grasp learning. CoRR, 2024

  8. [16]

    3d equivariant visuomotor policy learning via spherical projection

    Hu, B., Wang, D., Klee, D., Tian, H., Zhu, X., Huang, H., Platt, R., and Walters, R. 3d equivariant visuomotor policy learning via spherical projection. arXiv preprint arXiv:2505.16969, 2025

  9. [17]

    Edge grasp network: A graph-based se (3)-invariant approach to grasp detection

    Huang, H., Wang, D., Zhu, X., Walters, R., and Platt, R. Edge grasp network: A graph-based se (3)-invariant approach to grasp detection. arXiv preprint arXiv:2211.00191, 2022

  10. [18]

    L., Wang, D., Zhu, X., Platt, R., and Walters, R

    Huang, H., Howell, O. L., Wang, D., Zhu, X., Platt, R., and Walters, R. Fourier transporter: Bi-equivariant robotic manipulation in 3d. In The Twelfth International Conference on Learning Representations, 2024 a

  11. [19]

    IMAGINATION POLICY : Using generative point cloud models for learning manipulation policies

    Huang, H., Schmeckpeper, K., Wang, D., Biza, O., Qian, Y., Liu, H., Jia, M., Platt, R., and Walters, R. IMAGINATION POLICY : Using generative point cloud models for learning manipulation policies. In 8th Annual Conference on Robot Learning, 2024 b . URL https://openreview.net/...

  12. [20]

    Leveraging symmetries in pick and place

    Huang, H., Wang, D., Tangri, A., Walters, R., and Platt, R. Leveraging symmetries in pick and place. The International Journal of Robotics Research, 43 0 (4): 0 550--571, 2024 c

  13. [21]

    Planning with diffusion for flexible behavior synthesis

    Janner, M., Du, Y., Tenenbaum, J., and Levine, S. Planning with diffusion for flexible behavior synthesis. In International Conference on Machine Learning, pp.\ 9902--9915. PMLR, 2022

  14. [22]

    Seil: simulation-augmented equivariant imitation learning

    Jia, M., Wang, D., Su, G., Klee, D., Zhu, X., Walters, R., and Platt, R. Seil: simulation-augmented equivariant imitation learning. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pp.\ 1845--1851. IEEE, 2023

  15. [23]

    Se (3) diffusion model-based point cloud registration for robust 6d object pose estimation

    Jiang, H., Salzmann, M., Dang, Z., Xie, J., and Yang, J. Se (3) diffusion model-based point cloud registration for robust 6d object pose estimation. Advances in Neural Information Processing Systems, 36: 0 21285--21297, 2023

  16. [24]

    Grasping with chopsticks: Combating covariate shift in model-free imitation learning for fine manipulation

    Ke, L., Wang, J., Bhattacharjee, T., Boots, B., and Srinivasa, S. Grasping with chopsticks: Combating covariate shift in model-free imitation learning for fine manipulation. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pp.\ 6185--6191. IEEE, 2021

  17. [25]

    3d diffuser actor: Policy diffusion with 3d scene representations

    Ke, T.-W., Gkanatsios, N., and Fragkiadaki, K. 3d diffuser actor: Policy diffusion with 3d scene representations. arXiv preprint arXiv:2402.10885, 2024

  18. [26]

    Image to sphere: Learning equivariant features for efficient pose prediction

    Klee, D., Biza, O., Platt, R., and Walters, R. Image to sphere: Learning equivariant features for efficient pose prediction. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=_2bDpAtr7PI

  19. [27]

    S., Arora, E., and Platt, R

    Kohler, C., Srikanth, A. S., Arora, E., and Platt, R. Symmetric models for visual force policy learning. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pp.\ 3101--3107, 2024. doi:10.1109/ICRA57147.2024.10610728

  20. [28]

    Equivariant flows: exact likelihood generative learning for symmetric densities

    K \"o hler, J., Klein, L., and No \'e , F. Equivariant flows: exact likelihood generative learning for symmetric densities. In International conference on machine learning, pp.\ 5361--5370. PMLR, 2020

  21. [29]

    Z., and Gershman, S

    Lai, L., Huang, A. Z., and Gershman, S. J. Action chunking as policy compression. 2022

  22. [30]

    and Smidt, T

    Liao, Y.-L. and Smidt, T. Equiformer: Equivariant graph attention transformer for 3d atomistic graphs. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=KwmPfARgOTD

  23. [31]

    M., Das, A., and Smidt, T

    Liao, Y.-L., Wood, B. M., Das, A., and Smidt, T. Equiformerv2: Improved equivariant transformer for scaling to higher-degree representations. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=mCOBKZmrzD

  24. [32]

    Continual vision-based reinforcement learning with group symmetries

    Liu, S., Xu, M., Huang, P., Zhang, X., Liu, Y., Oguchi, K., and Zhao, D. Continual vision-based reinforcement learning with group symmetries. In Tan, J., Toussaint, M., and Darvish, K. (eds.), Proceedings of The 7th Conference on Robot Learning, volume 229 of Proceedings of Ma...

  25. [33]

    Rdt-1b: a diffusion foundation model for bimanual manipulation, 2024

    Liu, S., Wu, L., Li, B., Tan, H., Chen, H., Wang, Z., Xu, K., Su, H., and Zhu, J. Rdt-1b: a diffusion foundation model for bimanual manipulation, 2024. URL https://arxiv.org/abs/2410.07864

  26. [34]

    Luo, S., Chen, T., and Krishnapriyan, A. S. Enabling efficient equivariant operations in the fourier basis via gaunt tensor products. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=mhyQXJ6JsK

  27. [35]

    What matters in learning from offline human demonstrations for robot manipulation

    Mandlekar, A., Xu, D., Wong, J., Nasiriany, S., Wang, C., Kulkarni, R., Fei-Fei, L., Savarese, S., Zhu, Y., and Mart \' n-Mart \' n, R. What matters in learning from offline human demonstrations for robot manipulation. arXiv preprint arXiv:2108.03298, 2021

  28. [36]

    Mimicgen: A data generation system for scalable robot learning using human demonstrations

    Mandlekar, A., Nasiriany, S., Wen, B., Akinola, I., Narang, Y., Fan, L., Zhu, Y., and Fox, D. Mimicgen: A data generation system for scalable robot learning using human demonstrations. In Conference on Robot Learning, pp.\ 1820--1864. PMLR, 2023

  29. [37]

    K., Geiger, M., Smidt, T

    Miller, B. K., Geiger, M., Smidt, T. E., and No \'e , F. Relevance of rotationally equivariant convolutions for predicting molecular properties. arXiv preprint arXiv:2008.08461, 2020

  30. [38]

    6-dof graspnet: Variational grasp generation for object manipulation

    Mousavian, A., Eppner, C., and Fox, D. 6-dof graspnet: Variational grasp generation for object manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019

  31. [39]

    and Zitnick, C

    Passaro, S. and Zitnick, C. L. Reducing so (3) convolutions to so (2) for efficient equivariant gnns. In International Conference on Machine Learning, pp.\ 27420--27438. PMLR, 2023

  32. [40]

    V., Tan, S

    Pearce, T., Rashid, T., Kanervisto, A., Bignell, D., Sun, M., Georgescu, R., Macua, S. V., Tan, S. Z., Momennejad, I., Hofmann, K., et al. Imitating human behaviour with diffusion models. In The Eleventh International Conference on Learning Representations, 2023

  33. [41]

    Film: Visual reasoning with a general conditioning layer

    Perez, E., Strub, F., De Vries, H., Dumoulin, V., and Courville, A. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018

  34. [42]

    Diffusion-edfs: Bi-equivariant denoising generative modeling on se (3) for visual robotic manipulation

    Ryu, H., Kim, J., An, H., Chang, J., Seo, J., Kim, T., Kim, Y., Hwang, C., Choi, J., and Horowitz, R. Diffusion-edfs: Bi-equivariant denoising generative modeling on se (3) for visual robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Patter...

  35. [43]

    u ndung der theorie der gruppencharaktere. In Sitzungsberichte der K \

    Schur, I. Neue begr \"u ndung der theorie der gruppencharaktere. In Sitzungsberichte der K \"o niglich Preu ischen Akademie der Wissenschaften zu Berlin: Jahrgang 1905; Erster Halbband Januar bis Juni , pp.\ 406--432. Verlag der K \"o niglichen Akademie der Wissenschaften, 1905

  36. [44]

    Serre, J.-P. et al. Linear representations of finite groups, volume 42. Springer, 1977

  37. [45]

    B., Rodriguez, A., Agrawal, P., and Sitzmann, V

    Simeonov, A., Du, Y., Tagliasacchi, A., Tenenbaum, J. B., Rodriguez, A., Agrawal, P., and Sitzmann, V. Neural descriptor fields: Se (3)-equivariant object representations for manipulation. In 2022 International Conference on Robotics and Automation (ICRA), pp.\ 6394--6400. IEEE, 2022

  38. [46]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., and Ganguli, S. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pp.\ 2256--2265. PMLR, 2015

  39. [47]

    Denoising diffusion implicit models

    Song, J., Meng, C., and Ermon, S. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021

  40. [48]

    Et-seed: Efficient trajectory-level se (3) equivariant diffusion policy

    Tie, C., Chen, Y., Wu, R., Dong, B., Li, Z., Gao, C., and Dong, H. Et-seed: Efficient trajectory-level se (3) equivariant diffusion policy. arXiv preprint arXiv:2411.03990, 2024

  41. [49]

    Mujoco: A physics engine for model-based control

    Todorov, E., Erez, T., and Tassa, Y. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ international conference on intelligent robots and systems, pp.\ 5026--5033. IEEE, 2012

  42. [50]

    Se (3)-diffusionfields: Learning smooth cost functions for joint grasp and motion optimization through diffusion

    Urain, J., Funk, N., Peters, J., and Chalvatzaki, G. Se (3)-diffusionfields: Learning smooth cost functions for joint grasp and motion optimization through diffusion. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pp.\ 5923--5930. IEEE, 2023

  43. [51]

    Mdp homomorphic networks: Group symmetries in reinforcement learning

    Van der Pol, E., Worrall, D., van Hoof, H., Oliehoek, F., and Welling, M. Mdp homomorphic networks: Group symmetries in reinforcement learning. Advances in Neural Information Processing Systems, 33: 0 4199--4210, 2020

  44. [52]

    Rise: 3d perception makes real-world robot imitation simple and effective

    Wang, C., Fang, H., Fang, H.-S., and Lu, C. Rise: 3d perception makes real-world robot imitation simple and effective. arXiv preprint arXiv:2404.12281, 2024 a

  45. [53]

    Equivariant \ q\ learning in spatial action spaces

    Wang, D., Walters, R., Zhu, X., and Platt, R. Equivariant \ q\ learning in spatial action spaces. In 5th Annual Conference on Robot Learning, 2021. URL https://openreview.net/forum?id=IScz42A3iCI

  46. [54]

    On-robot learning with equivariant models

    Wang, D., Jia, M., Zhu, X., Walters, R., and Platt, R. On-robot learning with equivariant models. In 6th Annual Conference on Robot Learning, 2022 a

  47. [55]

    SO(2) -equivariant reinforcement learning

    Wang, D., Walters, R., and Platt, R. SO(2) -equivariant reinforcement learning. In International Conference on Learning Representations, 2022 b

  48. [56]

    Equivariant diffusion policy

    Wang, D., Hart, S., Surovik, D., Kelestemur, T., Huang, H., Zhao, H., Yeatman, M., Wang, J., Walters, R., and Platt, R. Equivariant diffusion policy. In 8th Annual Conference on Robot Learning, 2024 b

  49. [57]

    Y., Jia, M., Su, G., Platt, R., and Walters, R

    Wang, D., Zhu, X., Park, J. Y., Jia, M., Su, G., Platt, R., and Walters, R. A general theory of correct, incorrect, and extrinsic equivariance. Advances in Neural Information Processing Systems, 36, 2024 c

  50. [58]

    GELLO : A general, low-cost, and intuitive teleoperation framework for robot manipulators

    Wu, P., Shentu, F., Lin, X., and Abbeel, P. GELLO : A general, low-cost, and intuitive teleoperation framework for robot manipulators. In Towards Generalist Robots: Learning Paradigms for Scalable Skill Acquisition @ CoRL2023, 2023. URL https://openreview.net/forum?id=sseGcw79Zh

  51. [59]

    Geodiff: A geometric diffusion model for molecular conformation generation

    Xu, M., Yu, L., Song, Y., Shi, C., Ermon, S., and Tang, J. Geodiff: A geometric diffusion model for molecular conformation generation. In International Conference on Learning Representations, 2022

  52. [60]

    Equibot: Sim (3)-equivariant diffusion policy for generalizable and data efficient learning

    Yang, J., Cao, Z., Deng, C., Antonova, R., Song, S., and Bohg, J. Equibot: Sim (3)-equivariant diffusion policy for generalizable and data efficient learning. In 8th Annual Conference on Robot Learning, 2024 a

  53. [61]

    Equivact: Sim (3)-equivariant visuomotor policies beyond rigid object manipulation

    Yang, J., Deng, C., Wu, J., Antonova, R., Guibas, L., and Bohg, J. Equivact: Sim (3)-equivariant visuomotor policies beyond rigid object manipulation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pp.\ 9249--9255. IEEE, 2024 b

  54. [62]

    L., De Bortoli, V., Mathieu, E., Doucet, A., Barzilay, R., and Jaakkola, T

    Yim, J., Trippe, B. L., De Bortoli, V., Mathieu, E., Doucet, A., Barzilay, R., and Jaakkola, T. Se (3) diffusion model with application to protein backbone generation. In Proceedings of the 40th International Conference on Machine Learning, pp.\ 40001--40039, 2023

  55. [63]

    3d diffusion policy

    Ze, Y., Zhang, G., Zhang, K., Hu, C., Wang, M., and Xu, H. 3d diffusion policy. arXiv preprint arXiv:2403.03954, 2024

  56. [64]

    R., Bauza, M., Ma, D., Taylor, O., Liu, M., Romo, E., et al

    Zeng, A., Song, S., Yu, K.-T., Donlon, E., Hogan, F. R., Bauza, M., Ma, D., Taylor, O., Liu, M., Romo, E., et al. Robotic pick-and-place of novel objects in clutter with multi-affordance grasping and cross-domain image matching. The International Journal of Robotics Research, ...

  57. [65]

    H., and Koltun, V

    Zhao, H., Jiang, L., Jia, J., Torr, P. H., and Koltun, V. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 16259--16268, 2021

  58. [66]

    Hierarchical equivariant policy via frame transf

    Zhao, H., Wang, D., Zhu, Y., Zhu, X., Howell, O., Zhao, L., Qian, Y., Walters, R., and Platt, R. Hierarchical equivariant policy via frame transf. arXiv preprint arXiv:2502.05728, 2025

  59. [67]

    Zhao, L., Zhu, X., Kong, L., Walters, R., and Wong, L. L. Integrating symmetry into differentiable planning with steerable convolutions. In The Eleventh International Conference on Learning Representations, 2023 a

  60. [68]

    Z., Kumar, V., Levine, S., and Finn, C

    Zhao, T. Z., Kumar, V., Levine, S., and Finn, C. Learning fine-grained bimanual manipulation with low-cost hardware. arXiv preprint arXiv:2304.13705, 2023 b

  61. [69]

    Sample efficient grasp learning using equivariant models

    Zhu, X., Wang, D., Biza, O., Su, G., Walters, R., and Platt, R. Sample efficient grasp learning using equivariant models. Proceedings of Robotics: Science and Systems (RSS), 2022 a

  62. [70]

    On robot grasp learning using equivariant models

    Zhu, X., Wang, D., Su, G., Biza, O., Walters, R., and Platt, R. On robot grasp learning using equivariant models. Autonomous Robots, 2023

  63. [71]

    Coarse-to-fine 3d keyframe transporter

    Zhu, X., Klee, D., Wang, D., Hu, B., Huang, H., Tangri, A., Walters, R., and Platt, R. Coarse-to-fine 3d keyframe transporter. arXiv preprint arXiv:2502.01773, 2025 a

  64. [72]

    Equact: An se(3)-equivariant multi-task transformer for open-loop robotic manipulation, 2025 b

    Zhu, X., Qi, Y., Zhu, Y., Walters, R., and Platt, R. Equact: An se(3)-equivariant multi-task transformer for open-loop robotic manipulation, 2025 b . URL https://arxiv.org/abs/2505.21351

  65. [73]

    Viola: Object-centric imitation learning for vision-based robot manipulation

    Zhu, Y., Joshi, A., Stone, P., and Zhu, Y. Viola: Object-centric imitation learning for vision-based robot manipulation. In 6th Annual Conference on Robot Learning, 2022 b

  66. [74]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.