Lie Diffuser Actor for Robotic Manipulation

• Lie Diffuser Actor (LDA) is a diffusion-based vision-language-action policy that operates directly on the SE(3) manifold, overcoming the Euclidean fallacy in trajectory-level manipulation.
• It employs intrinsic diffusion by injecting Gaussian noise in the Lie algebra (𝔰𝔢(3)) and using the exponential map, thereby preserving manifold structure and ensuring coordinate-frame equivariance.
• Empirical results on CALVIN benchmarks show that LDA generates smoother, geodesic trajectories and improves task-chain success rates compared to traditional Euclidean diffusion methods.

Searching arXiv for the cited papers to ground the article in current metadata and related context. arXiv search query: (Chuang et al., 1 Jun 2026) Lie Diffuser Actor (LDA) is a diffusion-based Vision-Language-Action policy for robotic manipulation that operates intrinsically on $SE(3)$ rather than flattening rigid poses into a Euclidean vector such as $\mathbb{R}^{12}$. In the formulation reported for trajectory-level manipulation, the policy conditions on RGB-D observations $V=\{I_1,\dots,I_K\}$ from $K$ cameras and a natural-language instruction $C$, and generates an end-effector pose trajectory $\mathbf{g}=(g_1,\dots,g_H)$, with $g_h\in SE(3)$, as a conditional generative model $p_\theta(\mathbf{g}\mid V,C)$. The framework is motivated by what it terms the Euclidean Fallacy: treating $SE(3)$, a Lie group and Riemannian manifold, as a flat vector space for forward diffusion and denoising. The reported consequences are manifold drift, broken equivariance under coordinate transformations, and non-geodesic trajectories with excessive kinematic cost. On CALVIN ABC$\rightarrow$D, LDA improves average task length from $\mathbb{R}^{12}$0 to $\mathbb{R}^{12}$1 $\mathbb{R}^{12}$2, and the same study reports better or equal success rates on a majority of real-robot tasks (Chuang et al., 1 Jun 2026).

1. Problem setting and the Euclidean Fallacy

LDA is formulated for trajectory-level manipulation in which the action space is a sequence of rigid end-effector poses. Prior diffusion-based VLA policies considered in the same line of work encode visual input into a 3D representation, run a transformer over trajectory tokens with diffusion time as input, predict additive noise in a Euclidean representation of poses such as a $\mathbb{R}^{12}$3-D vector per $\mathbb{R}^{12}$4 pose, and sample in $\mathbb{R}^{12}$5. LDA rejects that construction on geometric grounds and instead treats the rigid-motion state as an element of $\mathbb{R}^{12}$6 throughout the diffusion process (Chuang et al., 1 Jun 2026).

The Euclidean Fallacy is defined as treating $\mathbb{R}^{12}$7 as a flat vector space for noise injection and denoising, and then projecting back. In the standard Euclidean forward diffusion on a single pose,

$\mathbb{R}^{12}$8

the state $\mathbb{R}^{12}$9 lives in a flattened Euclidean pose space. By contrast,

$V=\{I_1,\dots,I_K\}$0

Adding Gaussian noise to the $V=\{I_1,\dots,I_K\}$1 entries of a rotation matrix almost surely produces a non-orthogonal matrix.

Three consequences are identified. First, manifold drift arises because intermediate states leave $V=\{I_1,\dots,I_K\}$2. With Euclidean diffusion,

$V=\{I_1,\dots,I_K\}$3

and

$V=\{I_1,\dots,I_K\}$4

The deviation

$V=\{I_1,\dots,I_K\}$5

is almost surely nonzero for $V=\{I_1,\dots,I_K\}$6, so $V=\{I_1,\dots,I_K\}$7 with probability zero. Second, broken equivariance under coordinate transformations follows because additive Gaussian noise in $V=\{I_1,\dots,I_K\}$8 does not transform covariantly under rigid changes of frame. Third, non-geodesic, kinematically suboptimal trajectories arise because linear interpolation in Euclidean pose coordinates does not follow screw motions and generically does not remain on $V=\{I_1,\dots,I_K\}$9.

A common misconception is that these issues can be reduced to a harmless parameterization choice. In the LDA account, they are structural: Euclidean diffusion forces the model to learn to repair invalid rotations through normalization or SVD-like projection rather than focusing solely on manipulation behavior.

2. Geometric foundations on $K$0 and $K$1

The rigid-motion group is represented by

$K$2

with group product

$K$3

This makes $K$4 a $K$5-D Lie group and a curved Riemannian manifold. LDA moves the stochastic component of diffusion to the tangent space at the identity, namely the Lie algebra $K$6, whose elements are twists $K$7 (Chuang et al., 1 Jun 2026).

The hat representation is

$K$8

Because $K$9 is a flat vector space, it is the natural domain for Gaussian noise. The exponential map sends twists back to valid rigid motions: $C$0 where

$C$1

and

$C$2

Geometrically, $C$3 is the pose obtained by integrating a constant twist for unit time, namely a screw motion. This is the basis of what the method calls manifold correctness: for any $C$4, $C$5, and $C$6 is closed under composition.

3. Intrinsic diffusion and tangent-space score matching

The central redefinition in LDA is to replace additive Euclidean noising with multiplicative noising on the group: $C$7 Noise is therefore applied through group composition, the random variable lives in $C$8, the network predicts scores in tangent space, and updates return to the manifold through the exponential map. The corresponding continuous-time forward process is a left-invariant Stratonovich SDE,

$C$9

with $\mathbf{g}=(g_1,\dots,g_H)$0 an orthonormal basis of $\mathbf{g}=(g_1,\dots,g_H)$1 and $\mathbf{g}=(g_1,\dots,g_H)$2 independent Wiener processes (Chuang et al., 1 Jun 2026).

For generation, the reverse-time dynamics are written as

$\mathbf{g}=(g_1,\dots,g_H)$3

where $\mathbf{g}=(g_1,\dots,g_H)$4 approximates the score $\mathbf{g}=(g_1,\dots,g_H)$5. A discrete reverse update takes the form

$\mathbf{g}=(g_1,\dots,g_H)$6

Every step remains in $\mathbf{g}=(g_1,\dots,g_H)$7 by construction.

Training uses denoising score matching in tangent space. Given a clean trajectory $\mathbf{g}=(g_1,\dots,g_H)$8, one samples $\mathbf{g}=(g_1,\dots,g_H)$9, twists $g_h\in SE(3)$0, and noisy states

$g_h\in SE(3)$1

The core loss is

$g_h\in SE(3)$2

The full objective is

$g_h\in SE(3)$3

where $g_h\in SE(3)$4 is an MSE auxiliary loss on translations and $g_h\in SE(3)$5 is a BCE loss for gripper open/close.

4. Network architecture and generation pipeline

LDA builds on 3D Diffuser Actor but changes three components: the geometric context encoder, the action head, and the loss. The reported pipeline begins from RGB-D observations and a language instruction, projects depth images into a unified point cloud in robot or world coordinates using camera intrinsics and extrinsics, and encodes the geometry through a graph attention transformer and sparse 3D convolution to produce geometric features. The instruction is encoded with a CLIP text encoder, and cross-attention fuses geometric and language features into a multimodal context representation (Chuang et al., 1 Jun 2026).

At each diffusion step, a transformer processes the current noisy trajectory, a time embedding, and the multimodal context via cross-attention. Pose tokenization uses translational embeddings from the translation vector, axis-angle encoding of the rotation, and a gripper-state embedding. Self-attention operates over the $g_h\in SE(3)$6 waypoints, while prediction heads output a tangent-space twist

$g_h\in SE(3)$7

for each waypoint and a binary gripper action.

Sampling mirrors standard score-based diffusion, but on $g_h\in SE(3)$8. Initialization uses a noisy trajectory, and for $g_h\in SE(3)$9, the model conditions on the noisy poses, time embedding, and context, predicts $p_\theta(\mathbf{g}\mid V,C)$0, samples Gaussian noise in $p_\theta(\mathbf{g}\mid V,C)$1, and updates each waypoint through the group exponential: $p_\theta(\mathbf{g}\mid V,C)$2 Conditioning on vision and language is injected at every denoising step.

This makes LDA more than a replacement of the output head. The method changes the state space, the forward noising process, the reverse update, and the denoising target.

5. Theoretical properties

The framework states three principal guarantees. First, manifold drift is eliminated by construction. Proposition 4.1 asserts that the left-invariant SDE keeps $p_\theta(\mathbf{g}\mid V,C)$3 for all $p_\theta(\mathbf{g}\mid V,C)$4 almost surely. Since each update is $p_\theta(\mathbf{g}\mid V,C)$5 with $p_\theta(\mathbf{g}\mid V,C)$6, closure of the group suffices to keep all iterates valid (Chuang et al., 1 Jun 2026).

Second, coordinate-frame equivariance is expressed as a left-invariance property. For optimal score $p_\theta(\mathbf{g}\mid V,C)$7 and any $p_\theta(\mathbf{g}\mid V,C)$8,

$p_\theta(\mathbf{g}\mid V,C)$9

where the adjoint action for $SE(3)$0 on a twist $SE(3)$1 is

$SE(3)$2

In the reported interpretation, if the workspace and initial conditions are transformed by $SE(3)$3, the optimal score transforms accordingly, so the policy is intrinsic to the task geometry rather than to arbitrary frame conventions.

Third, geodesic trajectories arise in the probability-flow analysis. The deterministic probability flow ODE is

$SE(3)$4

Proposition 4.3 states that when the score is constant along the trajectory,

$SE(3)$5

the flow generates geodesics on $SE(3)$6 under the bi-invariant metric, with solution

$SE(3)$7

These curves are screw motions with constant angular and linear velocities. In practice the score varies, but the paper argues that the intrinsic formulation biases the reverse dynamics toward screw-like behavior and lower kinematic cost.

A corresponding misconception is that post-hoc projection of Euclidean outputs to $SE(3)$8 is equivalent to intrinsic diffusion. The reported analysis rejects that equivalence: projection is inference-only, induces training-inference mismatch, can amplify small errors, and alters denoising dynamics in a way training never sees.

6. Empirical results and diagnostic analyses

The main benchmark is CALVIN, a long-horizon language-conditioned manipulation benchmark with $SE(3)$9 skills and four environments $\rightarrow$0. In the ABC$\rightarrow$1D setting, training uses $\rightarrow$2 and tests zero-shot on unseen $\rightarrow$3; in ABCD$\rightarrow$4D, training uses all four environments and tests on $\rightarrow$5. Reported metrics are success rates $\rightarrow$6 for chains of $\rightarrow$7 to $\rightarrow$8 tasks and average task-chain length (Chuang et al., 1 Jun 2026).

On CALVIN ABC$\rightarrow$9D, 3D Diffuser Actor at $\mathbb{R}^{12}$00k iterations attains average length $\mathbb{R}^{12}$01, whereas full LDA at $\mathbb{R}^{12}$02k attains $\mathbb{R}^{12}$03, corresponding to $\mathbb{R}^{12}$04. In ABCD$\mathbb{R}^{12}$05D, the Euclidean baseline attains $\mathbb{R}^{12}$06, while full LDA at $\mathbb{R}^{12}$07k attains $\mathbb{R}^{12}$08. The ablations reported in the same study indicate that both the GAT encoder and Lie diffusion improve performance independently, and that the combined model performs best.

Real-robot evaluation uses four tasks with $\mathbb{R}^{12}$09 trials each: Move Doll Platform, Put Block in Box, Sort Blocks, and Stack Cups. The reported qualitative differences are task dependent. For Move Doll Platform and Stack Cups, LDA produces smoother $\mathbb{R}^{12}$10-DOF screw motions, more consistent orientation control, and more stable insertion behavior, whereas the Euclidean baseline shows segmented movement and orientation corrections. For Sort Blocks, LDA reportedly better discriminates foreground and background and maintains collinearity. For Put Block in Box, where translation dominates over strong rotation-translation coupling, LDA is reported as comparable to the baseline.

The diffusion-manifold analysis is central to the empirical argument. Rotations visualized as quaternions on $\mathbb{R}^{12}$11 remain on the unit sphere under LDA and leave it under Euclidean diffusion. Orthogonality error $\mathbb{R}^{12}$12, determinant error $\mathbb{R}^{12}$13, and quaternion norm stay at approximately $\mathbb{R}^{12}$14 under LDA and show large violations under Euclidean diffusion, including quaternion norms in approximately $\mathbb{R}^{12}$15. The study also defines geodesic jitter as

$\mathbb{R}^{12}$16

measured in degrees for successive predictions of the final clean action. Euclidean jitter is reported as up to approximately $\mathbb{R}^{12}$17 higher than LDA, especially early in denoising.

Cross-architecture validation is performed by adding score-matching heads to OpenVLA-OFT on LIBERO Long. The baseline MLP regression head reports $\mathbb{R}^{12}$18 success, a Euclidean score-matching head reports $\mathbb{R}^{12}$19, and a Lie score-matching head reports $\mathbb{R}^{12}$20. Because the perceptual backbone is unchanged, the reported improvement is presented as evidence that Lie score matching is architecture-agnostic.

7. Relation to adjacent diffusion literature, nomenclature, limitations, and outlook

The label LDA is not entirely stable across recent diffusion literature. In the long-context diffusion-LLM study “LongLLaDA: Unlocking Long Context Capabilities in Diffusion LLMs,” the discussion describes LongLLaDA as a training-free length-extrapolation method applied on top of existing diffusion LLMs such as LLaDA. That work states that LongLLaDA does not modify the underlying diffusion architecture, the score network, or any Lie group structure; all changes are at the level of RoPE positional encoding at inference time, through NTK-aware scaling of the rotary base (Liu et al., 17 Jun 2025).

This establishes a useful contrast. In the robotics meaning of Lie Diffuser Actor, the intervention is geometric and intrinsic: the state lives on $\mathbb{R}^{12}$21, noise is injected in $\mathbb{R}^{12}$22, the score is learned in tangent space, and updates return to the manifold through the exponential map. In LongLLaDA, by contrast, the intervention is an inference-stage re-parameterization of RoPE for long-context extrapolation in a transformer denoiser. A plausible implication is that the acronym is overloaded across domains and should not be treated as denoting a single method family.

The limitations reported for robotic LDA are correspondingly specific. Operating on $\mathbb{R}^{12}$23 requires exponential and, at times, logarithm map computations per waypoint per step, though the reported implementation uses closed-form Rodrigues and left-Jacobian formulas that are described as relatively cheap on GPU. The formulation depends on accurate $\mathbb{R}^{12}$24 and sometimes $\mathbb{R}^{12}$25 implementations, large twists may require care for numerical stability, and extension beyond $\mathbb{R}^{12}$26 to more complex configuration spaces remains future work. Residual failures are attributed not to geometry but to perception errors or insufficient demonstrations.

The broader conceptual implication is a shift from extrinsic to intrinsic generative modeling for robot action. Rather than using “pose in $\mathbb{R}^{12}$27 + additive Gaussian noise + projection,” LDA uses “pose in $\mathbb{R}^{12}$28, noise in $\mathbb{R}^{12}$29, exponential-map retraction.” The reported future directions include combining intrinsic $\mathbb{R}^{12}$30 diffusion with more sophisticated equivariant perception, extending Lie diffusion to higher-level action spaces and other manifolds, and exploiting geodesic structure for energy-efficient and contact-aware motion.