Dynamics-Aware Motion Manifold Primitives

• The paper introduces a framework that learns a low-dimensional motion manifold from thousands of planned trajectories and refines them with a dynamics-aware generative model.
• It employs cubic Hermite spline parameterization and gated Gaussian RBFs to encode variable-length robot motions into compact latent codes via an autoencoder.
• Empirical results demonstrate DA-MMP achieving up to 60% success in dynamic tasks, outperforming traditional planning methods in real-world execution and novel target generalization.

Dynamics-Aware Motion Manifold Primitives (DA-MMP) are a framework for synthesizing coordinated, dynamically consistent robot motions, with explicit modeling of system dynamics and real-world execution feedback. The key innovation of DA-MMP is to learn a compact, variable-length low-dimensional manifold of motion trajectories, and then refine this manifold using dynamics-aware, goal-conditional generative models in latent space. This approach addresses the challenge of highly coordinated dynamic manipulation in robotics, particularly in settings where significant discrepancies exist between planned and real-world execution due to unmodeled control inaccuracies, timing uncertainties, and external disturbances such as aerodynamic drag or contact effects (Chu et al., 28 Sep 2025).

1. Motivation and Theoretical Foundations

Dynamic manipulation tasks, exemplified by throwing or rapid in-contact actions, require generating high-speed, temporally coordinated multi-joint trajectories. The traditional parameterizations (e.g., end-effector velocities or a small set of human demonstrations) are insufficiently expressive for these behaviors and cannot capture the rich space of dynamically feasible motions. Moreover, trajectories generated purely by motion planning—such as kinodynamic sampling-based planners—are not robust to real-world dynamics gaps, resulting in poor transfer from plan to execution.

DA-MMP is motivated by two core needs:

• Expressiveness: Modeling a manifold of high-dimensional, dynamically valid, variable-length motions efficiently.
• Dynamics Adaptation: Bridging the dynamics gap between simulated/planned and real world via learning-based correction using sparse real-world feedback, enabling robust generalization across unseen targets or tasks (Chu et al., 28 Sep 2025).

2. Variable-Length Motion Manifold Learning

DA-MMP generalizes prior Motion Manifold Primitive (MMP) frameworks by supporting variable-length trajectories and scaling to large numbers of planned motions. The core parameterization is as follows:

Each trajectory $q(t)$ of duration $L$ is reparameterized in normalized phase $s = t/L \in [0, 1]$ using:

• Cubic Hermite spline $\psi(s)$ fixed by start $\{q(0), \dot{q}(0)\}$ and end $\{q(1), \dot{q}(1)\}$ boundary conditions,
• Weighted sum of $K$ gated Gaussian radial basis functions $\boldsymbol{\phi}(s)$ with learned weights $\mathbf w$:

$$q(s; \mathbf w) = \psi(s) + \mathbf w^\top \boldsymbol{\phi}(s)$$

• Basis weights $L$0 and terminal state $L$1 plus trajectory length $L$2 are concatenated into a trajectory parameter vector $L$3.

The motion manifold is then built by autoencoding these $L$4-dimensional vectors $L$5 into a compact latent code $L$6 using a neural autoencoder. Training proceeds over a large corpus (e.g., $L$7k planned trajectories), optimizing the reconstruction loss: $L$8 This enables online generation of variable-length trajectories from the smooth function $L$9 for arbitrary $s = t/L \in [0, 1]$0 (Chu et al., 28 Sep 2025).

3. Dynamics-Aware Latent Generative Modeling

To correct for real-world dynamics discrepancies, DA-MMP introduces a conditional flow-matching model $s = t/L \in [0, 1]$1 in the learned low-dimensional latent space. The conditioning input $s = t/L \in [0, 1]$2 encodes real-world execution outcomes (e.g., measured landing point for throwing).

The flow-matching network interpolates between a Gaussian prior and empirical latent codes of real trials, training to match the ODE flow between samples: $s = t/L \in [0, 1]$3 where $s = t/L \in [0, 1]$4. This model is trained using only a small set of real executions (e.g., $s = t/L \in [0, 1]$5 trials), yielding exceptional data efficiency (Chu et al., 28 Sep 2025).

The outcome is a model that, given a new task condition $s = t/L \in [0, 1]$6 (target), can sample dynamically compensated motions by integrating $s = t/L \in [0, 1]$7, then decoding to full motion trajectories.

4. End-to-End Pipeline and Inference

The complete DA-MMP workflow comprises two stages:

Stage I: Manifold Construction and Autoencoding

1. Sample desired end-effector or object outcomes (e.g., release states for tossing).
2. Generate motion plans using kinodynamic RRT (with smoothing and double simulation filtering for plan robustness).
3. Parameterize each plan as $s = t/L \in [0, 1]$8 using the RBF-spline structure.
4. Train autoencoder on $s = t/L \in [0, 1]$9.

Stage II: Dynamics Learning and Conditioned Generation

1. Perform a small number of real-world executions; for each, record actual outcome $\psi(s)$0, encode corresponding $\psi(s)$1 to obtain $\psi(s)$2.
2. Train flow-matching model $\psi(s)$3 on pairs $\psi(s)$4.

At inference, given a new goal (e.g., target landing in the plane):

• Set $\psi(s)$5,
• Sample Gaussian prior latent $\psi(s)$6,
• Integrate flow ODE $\psi(s)$7 for $\psi(s)$8 to obtain $\psi(s)$9,
• Decode to $\{q(0), \dot{q}(0)\}$0 and reconstruct $\{q(0), \dot{q}(0)\}$1,
• Execute on the robot (Chu et al., 28 Sep 2025).

This pipeline leverages the expressive capacity of the learned manifold while allowing fast, dynamics-corrected online trajectory synthesis.

5. Empirical Evaluation and Comparison

DA-MMP has been evaluated on kinodynamic and dynamic manipulation benchmarks:

• On a real-world ring-tossing challenge, DA-MMP achieved a $\{q(0), \dot{q}(0)\}$2 real-world success rate, outperforming human experts ($\{q(0), \dot{q}(0)\}$3), motion planning ($\{q(0), \dot{q}(0)\}$4 for one attempt, $\{q(0), \dot{q}(0)\}$5 for two), and residual-style correction ($\{q(0), \dot{q}(0)\}$6). In simulation with drag, motion planning failed completely while DA-MMP maintained $\{q(0), \dot{q}(0)\}$7 (Chu et al., 28 Sep 2025).
• Generalizes to novel target locations beyond the training range, indicating that it learns a mapping from goal to trajectory rather than memorizing exemplars.

Ablative studies indicate the necessity of both the variable-length manifold model and dynamics-aware latent flow. Variant approaches lacking these mechanisms, or using classical MMPs trained on modest data without explicit dynamics feedback, exhibited much lower performance or failed to generate feasible trajectories (Chu et al., 28 Sep 2025).

6. Related Methodologies and Extensions

DA-MMP builds on several foundational methodologies:

• Motion Manifold Primitives (MMP): These define low-dimensional manifolds for motion generation but traditionally assume fixed-length trajectories and do not incorporate explicit dynamics feedback (Lee, 2024).
• Differentiable Motion Manifold Primitives (DMMP): Employ basis-function neural decoders, explicit kinodynamic constraint satisfaction, and flow-matching for fast planning but lack variable-length trajectory support and comprehensive dynamics adaptation (Lee, 2024).
• Geometry-aware DMPs: Generalize DMPs to Riemannian manifolds, enabling geometric and topological task constraint satisfaction (e.g., orientation, impedance), though not specifically targeting variable-length, dynamic or execution-conditioned tasks (Abu-Dakka et al., 2022, Vedove et al., 2024).
• MeshDMP-based DA-MMP for in-contact tasks: DA-MMP can be implemented in surface-constrained settings via MeshDMP, leveraging discrete-geometric operators for adaptation across mesh topologies, combined with impedance control for robust real-time in-contact execution (Vedove et al., 2024).

A representative synthesis is provided in the table below:

Approach	Variable-length	Dynamics-aware Feedback	Targeted Task Class
DA-MMP (Chu et al., 28 Sep 2025)	Yes	Yes	Free-space dynamic throws
DMMP (Lee, 2024)	No	Partial	Fast kinodynamic planning
Geometry-aware DMP (Abu-Dakka et al., 2022)	No	No	Manifold-constraint LfD
MeshDMP (Vedove et al., 2024)	Yes	Yes (with feedback)	In-contact, mesh-limited

7. Limitations and Future Directions

Key limitations include:

• Offline Data Requirements: Rich motion manifold learning requires extensive kinodynamic sampling and offline planning (e.g., $\{q(0), \dot{q}(0)\}$8k plans), potentially demanding in novel domains (Chu et al., 28 Sep 2025).
• Manual Engineering of Demonstration Space: The goal manifold or parameterization of desired outcomes must be explicitly designed.
• Constraint Enforcement: While DA-MMP can enforce dynamic feasibility via rejection-sampling or loss penalties, constraint satisfaction is not guaranteed, motivating future integration of learned feasibility classifiers (Lee, 2024).
• Extension to Contact-rich and Manifold Tasks: Uniform performance in highly varied, contact-rich, or non-Euclidean task spaces (e.g., on curved surfaces or with time-varying objectives) remains an area of active research.

A plausible implication is that DA-MMP general principles—decoupling manifold learning from dynamics adaptation, supporting variable-length trajectories, and using data-efficient latent-space flow models—will underpin scalable, robust motion generation frameworks for advanced robotic systems operating in uncertain and dynamic environments.

---

References:

• "DA-MMP: Learning Coordinated and Accurate Throwing with Dynamics-Aware Motion Manifold Primitives" (Chu et al., 28 Sep 2025)
• "Trajectory Manifold Optimization for Fast and Adaptive Kinodynamic Motion Planning" (Lee, 2024)
• "A Unified Formulation of Geometry-aware Dynamic Movement Primitives" (Abu-Dakka et al., 2022)
• "MeshDMP: Motion Planning on Discrete Manifolds using Dynamic Movement Primitives" (Vedove et al., 2024)