---
title: 'EquiContact: SE(3) Manipulation Framework'
url: https://www.emergentmind.com/topics/equicontact
type: topic
---

# EquiContact: SE(3) Manipulation Framework

Searching arXiv for the target paper and closely related work to ground the article.
arxiv_search(query="EquiContact 2507.10961 OR \"Hierarchical SE(3) Vision-to-Force Equivariant Policy\"", max_results=5, sort_by="submittedDate")
EquiContact is a hierarchical imitation-learning framework for contact-rich robotic manipulation that is designed to generalize across unseen spatial configurations by making the perception-to-force-control pipeline \(SE(3)\)-equivariant. It is formulated around peg-in-hole manipulation with \(1\mathrm{mm}\) clearance, where coarse global pose estimation is insufficient and robust success depends on combining global vision, localized reactive control, force feedback, and compliance. The framework consists of a high-level vision planner, Diffusion Equivariant Descriptor Field (Diff-EDF), and a low-level compliant visuomotor policy, Geometric Compliant ACT (G-CompACT), with geometric admittance control providing compliant execution [2507.10961].

## 1. Conceptual scope and problem setting

EquiContact addresses contact-rich manipulation rather than free-space reaching. In the peg-in-hole setting studied, the hole clearance is only \(1\mathrm{mm}\), while the high-level vision module produces approximate pose estimates with roughly \(5\)–\(15\mathrm{mm}\) translation error and up to tens of degrees of orientation error. This mismatch motivates a hierarchical design in which global perception supplies a coarse reference frame and local control performs the fine alignment and insertion [2507.10961].

The framework is organized around three stated principles: compliance, localized policies, and induced equivariance. Compliance is needed because contact events during insertion cannot be handled reliably by purely kinematic tracking. Localized policies are used because body-frame observations and actions can remain invariant under global task transformations. Induced equivariance is obtained by anchoring the local policy on a globally estimated reference frame so that world-frame behavior transforms consistently when the task configuration is translated or rotated [2507.10961].

A plausible implication is that EquiContact is less a single network architecture than a geometric systems design for contact-rich visuomotor control. This suggests that its main novelty lies in how perception, local sensing, action parameterization, and compliant control are coordinated, rather than in end-to-end monolithic policy learning.

## 2. Hierarchical architecture

EquiContact is composed of two modules. The high-level planner is Diff-EDF, which uses global point-cloud observations from external RGB-D cameras to estimate an approximate reference pose \(g_{EDF}\) for the task object. The low-level controller is G-CompACT, which uses localized observations—Geometrically Consistent Error Vectors, force-torque readings, and wrist-mounted RGB images—to predict relative end-effector actions and compliance gains [2507.10961].

The high-level planner is written as
\[
g_{EDF} = f_{\theta_1}(O^{scene}, O^{grasp}),
\]
where \(O^{scene}\) is the scene point cloud and \(O^{grasp}\) is the gripper point cloud expressed in the end-effector frame [2507.10961]. Diff-EDF is used only for coarse localization. The paper is explicit that its accuracy is insufficient for direct insertion in a \(1\mathrm{mm}\)-clearance task, so its output is used as a local reference frame for the low-level policy rather than as a final control target [2507.10961].

The low-level policy is written as
\[
a(k) = \pi_{\theta_2}(o(k)),
\qquad
o(k) \triangleq (\phi_G, F_e, I_w)(k),
\]
with chunked output
\[
a(k) \triangleq \{(g_{rel}(k+i), \bar{K}_p(k+i), \bar{K}_R(k+i))\}_{i=1}^{N}.
\]
Thus each action chunk contains relative poses in \(SE(3)\) and translational and rotational admittance gains [2507.10961].

The temporal composition rule is
\[
\bar{g}_d(k+i) = g(k) \cdot g_{rel}(k+i),
\]
so actions are defined in the end-effector frame and converted into spatial commands by composition with the current pose \(g(k)\) [2507.10961]. This action parameterization is central to the framework’s equivariance argument.

## 3. Geometric state representation and localized sensing

A core object in EquiContact is the Geometrically Consistent Error Vector, denoted \(\phi_G\). It is defined by
\[
\phi_G(g,g_{EDF}) =
\begin{bmatrix}
R^T (p - p_{EDF}) \\
(R_{EDF}^T R - R^T R_{EDF})^\vee
\end{bmatrix},
\]
where \(g=(p,R)\in SE(3)\) is the current end-effector pose and \(g_{EDF}=(p_{EDF},R_{EDF})\) is the reference pose predicted by Diff-EDF [2507.10961].

The translational component is a body-frame position error, and the rotational component is a body-frame orientation error. The paper emphasizes that \(\phi_G\) is left-invariant:
\[
g_l \circ \phi_G(g,g_{EDF}) = \phi_G(g_lg, g_lg_{EDF}) = \phi_G(g,g_{EDF}),
\]
for \(g_l \in SE(3)\) [2507.10961]. This left invariance is one of the exact geometric ingredients from which the equivariance of the full pipeline is built.

Localized sensing is equally important. The low-level policy uses force-torque readings \(F_e \in \mathbb{R}^6\) in the end-effector frame and two wrist-mounted RGB cameras \(I_w=\{I_{w,1},I_{w,2}\}\) [2507.10961]. Because both signals are end-effector-centered, global rigid transformations of the task do not directly alter their representation. The paper states an explicit assumption of left-invariant visual features:
\[
z(k) = \mu_\phi(g_l \circ I_w) = \mu_\phi(I_w),
\qquad \forall g_l \in SE(3),
\]
where \(z=\mu_\phi(I_w)\) is the wrist-image feature [2507.10961].

The paper also notes that this visual invariance is not enforced exactly by architecture or loss. Instead, it is induced approximately by the local perspective of the wrist cameras, the workspace design, and data support. This suggests that the geometric invariance of \(\phi_G\) and body-frame force is exact by construction, whereas the visual component is only approximate in practice.

## 4. Compliance and geometric admittance control

EquiContact does not directly map observations to low-level torques. Instead, G-CompACT predicts desired relative poses and compliance gains, which are executed through a geometric admittance controller based on geometric impedance control [2507.10961].

The desired end-effector dynamics are
\[
M \dot{V}^b + K_d V^b + \phi_G = F_e,
\]
where \(M \in \mathbb{R}^{6\times 6}\) is a symmetric positive-definite desired inertia matrix, \(K_d \in \mathbb{R}^{6\times 6}\) is a symmetric positive-definite damping matrix, \(V^b \in \mathbb{R}^6\) is the body-frame end-effector velocity, and \(F_e \in \mathbb{R}^6\) is the external wrench in the end-effector frame [2507.10961].

The elastic wrench is
\[
\phi_G =
\begin{bmatrix}
f_p \\
f_R
\end{bmatrix}
=
\begin{bmatrix}
R^T R_d K_p R_d^T (p - p_d) \\
(K_R R_d^T R - R^T R_d K_R)^\vee
\end{bmatrix},
\]
with symmetric positive translational and rotational stiffness matrices \(K_p, K_R \in \mathbb{R}^{3\times 3}\) [2507.10961].

The discrete-time controller is
\[
\begin{aligned}
V_d^b(k) &= V^b(k) + T_s \cdot M^{-1} \big( F_e(k) - \phi_G(k) - K_d V^b(k)\big), \\
\tilde{g}_d(k) &= g(k) \cdot \exp{(\hat{V}_d^b(k) \cdot T_s)}.
\end{aligned}
\]
The controller runs at \(200\mathrm{Hz}\) in ROS2 [2507.10961].

The policy outputs temporary gains \(\bar{K}_p,\bar{K}_R\) in the desired end-effector frame, and for the peg-in-hole tasks considered, diagonal matrices suffice [2507.10961]. This is a deliberate simplification relative to CompACT, where more general gain representations required full matrices and Cholesky factorization.

The experimental comparison shows that compliance is not merely an auxiliary refinement. ACT without geometric admittance control succeeds on only \(2/10\) in-distribution insertion trials, while ACT with geometric admittance control reaches \(20/20\) on the same setting [2507.10961]. This indicates that contact-rich insertion is not solved by equivariance alone; compliant execution is essential.

## 5. Equivariance structure

The paper’s central formal claim is that the EquiContact pipeline is \(SE(3)\)-equivariant from perception to force control under two assumptions: left-equivariance of Diff-EDF and left-invariance of wrist-image features [2507.10961].

The left-equivariance assumption for Diff-EDF is
\[
\begin{split}
g_{EDF} & = f_{\theta_1}(O^{ref}, O^{grasp}) \\
&\implies g_l \cdot g_{EDF} = f_{\theta_1}(g_l \circ O^{ref}, O^{grasp}),
\end{split}
\]
where \(O^{ref}\subset \mathcal{O}^{scene}\) is the point cloud of the target object [2507.10961].

Under the left-invariant feature assumption for wrist vision and the left invariance of \(\phi_G\) and \(F_e\), the paper states:

\[
a(k) = \pi_{\theta_2}(g_l \circ o(k)) = \pi_{\theta_2}(o(k)).
\]

This is Proposition 1, labeled “Left-invariance of G-CompACT” [2507.10961]. The low-level policy is therefore invariant in its local frame.

The world-frame desired pose remains equivariant because the invariant local action is composed with the transformed current pose. The paper states:

\[
(g_l g_d, K_p, K_R)(k) = \hat{\pi}_{\theta_2}(g_l \circ o(k)).
\]

This is Corollary 1, labeled “\(SE(3)\) left equivariance of G-CompACT” [2507.10961].

For the full pipeline, the policy \(h_\theta\) mapping current pose, reference pose, and force to elastic wrench is shown to satisfy a spatial-frame wrench equivariance law:
\[
h^s_\theta(g_l g, g_l g_{ref}, g_l \circ F_e)
=
Ad_{g_l^{-1}}^T h_\theta^s(g, g_{ref}, F_e).
\]
This elevates the result beyond pose-command equivariance to equivariance of interaction wrench behavior [2507.10961].

A plausible implication is that the framework operationalizes equivariance through coordinate design, signal localization, and composition rules, rather than requiring every learned module to be an exactly equivariant neural operator.

## 6. Training regime and data support

The two components are trained separately. Diff-EDF is trained from \(15\) demonstrations total: \(12\) on a flat platform and \(3\) on a tilted platform, with translational and rotational randomization and visual distractors [2507.10961]. G-CompACT is trained from \(86\) teleoperated insertion demonstrations collected on a fixed platform with known hole location [2507.10961].

Teleoperation uses a SpaceMouse for motion commands and keyboard switching among four predefined gain modes. Data is logged at \(30\mathrm{Hz}\), while teleoperation and robot control run at \(200\mathrm{Hz}\) [2507.10961]. The paper states that training follows the standard imitation-learning pipeline for ACT. It does not provide the full CVAE objective, optimizer details, or Diff-EDF diffusion loss [2507.10961].

A key training intervention is reference-pose noise injection for G-CompACT, because at inference the reference comes from imperfect Diff-EDF predictions. The noisy reference is defined by
\[
\begin{split}
\tilde{p}_{ref} &= p_{ref} + n_p,\\
\tilde{R}_{ref} &= R_{ref} \cdot \mathtt{euler2rotmat}(n_r),
\end{split}
\]
with
\[
n_p \sim \mathcal{U}(-0.02, 0.02),
\qquad
n_r \sim \mathcal{U}(-8^{\circ}, 8^\circ).
\]
Thus training includes \(2\mathrm{cm}\) translational and \(8^\circ\) rotational perturbations on the reference frame [2507.10961].

When the visual-invariance assumption broke down more severely, the paper augmented the data with \(20\) demonstrations containing visual distractors on the flat platform and \(20\) demonstrations on a \(30^\circ\) tilted platform [2507.10961]. This suggests that induced equivariance in the visual channel depends materially on the support of the demonstration set.

## 7. Empirical performance

The main real-world benchmark is peg-in-hole insertion under translated and tilted out-of-distribution platform configurations [2507.10961]. The comparison includes ACT without geometric admittance control, ACT with geometric admittance control, CompACT, and EquiContact.

The reported results are summarized below.

| Method | Setting | Success |
|---|---|---:|
| ACT w/o GAC | Flat Platform (In-Dist.) | \(2 / 10\) |
| ACT w/ GAC | Flat Platform (In-Dist.) | \(20 / 20\) |
| CompACT | Flat Platform (In-Dist.) | \(20 / 20\) |
| CompACT | Flat Platform (OOD) | \(0 / 10\) |
| EquiContact | Flat Platform (OOD) | \(20 / 20\) |
| EquiContact | Tilted Platform (\(30^\circ\), OOD) | \(19 / 20\) |

These results show three distinct phenomena. First, compliance is decisive for insertion, since ACT without GAC fails frequently while ACT with GAC succeeds in-distribution. Second, world-frame compliant imitation does not generalize spatially, since CompACT drops from \(20/20\) in-distribution to \(0/10\) out-of-distribution. Third, EquiContact preserves high performance under translated and tilted OOD configurations, consistent with the framework’s equivariance claims [2507.10961].

For the full pick-and-place pipeline, EquiContact achieves \(20/20\) on flat-platform OOD trials and \(18/20\) on \(30^\circ\)-tilted OOD trials [2507.10961]. The paper attributes the small degradation on the tilted case to error propagation across sequential stages, specifically one pick failure and one place failure [2507.10961].

The force-profile comparison further shows that CompACT yields lower contact forces than ACT with fixed gains, supporting the benefit of learned compliance modulation via force feedback [2507.10961].

## 8. Failure modes and limitations

The framework’s theoretical equivariance depends on assumptions that are not fully guaranteed in learned perception. The most important fragile component is the left-invariance of wrist-image features [2507.10961]. The paper reports degradation in the presence of visual distractors and larger tilt angles.

Using ground-truth reference frames and the base dataset, EquiContact achieves \(4/10\) under visual distractors and \(5/10\) on a \(45^\circ\)-tilted platform. After augmenting the data, these improve to \(9/10\) and \(9/10\), respectively [2507.10961]. This indicates that the visual channel does not inherit the same exact invariance as the geometric channels and requires data support to remain robust.

The paper also states several design assumptions. The peg is assumed upright. Right-equivariance is avoided by enforcing a consistent grasp orientation, because arbitrary transformations of the peg relative to the gripper caused slippage and would require continuous estimation of the object pose relative to the gripper [2507.10961]. External RGB-D cameras with calibrated extrinsics, wrist RGB cameras, and an end-effector force-torque sensor are assumed available [2507.10961].

Diff-EDF is not real-time, which is one reason the reactive low-level layer is necessary [2507.10961]. The system is hierarchical and modular rather than jointly end-to-end trained. A plausible implication is that EquiContact trades architectural simplicity for stronger geometric structure and clearer functional decomposition.

## 9. Relation to adjacent contact-centric work

EquiContact belongs to a line of work that treats contact as a structured signal rather than a purely latent nuisance. In robotic loco-manipulation, "OmniContact: Chaining Meta-Skills via Contact Flow for Generalizable Humanoid Loco-Manipulation" introduces Contact Flow as a compact representation consisting of key body trajectories and time-series binary contact signals, emphasizing explicit contact structure for compositional control [2606.26201]. The two systems differ substantially in setting and mechanism: OmniContact is aimed at long-horizon humanoid loco-manipulation and does not use explicit equivariant modeling, whereas EquiContact is centered on \(SE(3)\)-equivariant spatial generalization in contact-rich manipulation [2606.26201], [2507.10961].

This suggests a broader methodological split within contact-centric learning. One branch uses explicit contact representations for planning and skill composition, while another—exemplified by EquiContact—uses group-structured geometry, local sensing, and compliant control to make contact behavior spatially transferable.

## 10. Significance

EquiContact’s main contribution is the claim that robust spatial generalization in contact-rich manipulation can be achieved from a small number of demonstrations by structuring the entire pipeline around \(SE(3)\) geometry. The framework does not rely on exact pose estimation for insertion, nor on world-frame visuomotor mappings that must be relearned at each spatial configuration. Instead, it combines a coarse equivariant global planner, a localized low-level policy, body-frame force sensing, relative action composition, and compliant geometric control into a unified policy whose behavior transforms consistently across rigid scene transformations [2507.10961].

For contact-rich assembly, the framework provides an explicit recipe: estimate a global reference frame coarsely, represent local error geometrically in the end-effector frame, sense contact locally, output end-effector-frame relative actions and gains, and execute them through compliant control. The empirical results on out-of-distribution peg-in-hole insertion and pick-and-place indicate that this recipe can outperform world-frame imitation baselines by a large margin under limited-data conditions [2507.10961].

Source: https://www.emergentmind.com/topics/equicontact