DYNAMO: Dependency-Aware Deep Learning Framework for Articulated Assembly Motion Prediction

Understanding the motion of articulated mechanical assemblies from static geometry remains a core challenge in 3D perception and design automation. Prior work on everyday articulated objects such as doors and laptops typically assumes simplified kinematic structures or relies on joint annotations. However, in mechanical assemblies like gears, motion arises from geometric coupling, through meshing teeth or aligned axes, making it difficult for existing methods to reason about relational motion from geometry alone. To address this gap, we introduce MechBench, a benchmark dataset of 693 diverse synthetic gear assemblies with part-wise ground-truth motion trajectories. MechBench provides a structured setting to study coupled motion, where part dynamics are induced by contact and transmission rather than predefined joints. Building on this, we propose DYNAMO, a dependency-aware neural model that predicts per-part SE(3) motion trajectories directly from segmented CAD point clouds. Experiments show that DYNAMO outperforms strong baselines, achieving accurate and temporally consistent predictions across varied gear configurations. Together, MechBench and DYNAMO establish a novel systematic framework for data-driven learning of coupled mechanical motion in CAD assemblies.

DYNAMO is a dependency-aware neural network designed to predict how parts of a mechanical assembly move together. Starting from static CAD point clouds, it first extracts shape features for each part, then reasons about which parts are physically coupled, and finally predicts their motion over time. This three-stage process (geometry → coupling → motion) allows DYNAMO to capture both the local shape and the global structure of the assembly, producing realistic trajectories frame by frame.

To train and evaluate our model, we created MechBench, a dataset of 693 synthetic gear assemblies. It spans spur gears, bevels, racks, worms, and planetary systems, covering both simple pairs and complex multi-gear mechanisms. Each assembly comes with 3D meshes, segmentation masks, part-level point clouds, and precise motion annotations, making it a comprehensive testbed for studying coupled motion. Unlike previous datasets, MechBench explicitly encodes dependency-driven motion, where one part’s rotation induces another’s, offering a unique challenge for learning-based methods.

1. Spur Gear
2. Helical Spur Gear
3. Double Helical Spur Gear
4. Rack and Pinion
5. Helical Rack and Pinion
6. Double Helical Rack and Pinion
7. Internal Involute Gear
8. Helical Internal Involute Gear
9. Double Helical Internal Involute Gear
10. Bevel Gear
11. Helical Bevel Gear
12. Work Gear
13. Screw Gear

DYNAMO predicts both rotational and translational motions with high fidelity across a wide range of assemblies. It handles synchronized gear rotations, linear motion in rack-and-pinion systems, and even the transfer of motion across orthogonal in screw and bevel gears. In each case, predictions stay tightly aligned with the ground truth and remain smooth over time, preserving contact constraints and correct rotational direction.

Remarkably, DYNAMO learns coupling behavior directly from geometry, without explicit supervision of axes, gear ratios, or kinematic rules. For instance, it infers that smaller gears rotate faster when meshed with larger ones, effectively capturing gear ratios from part geometry alone. This emergent ability highlights the model’s strength in internalizing physical constraints and producing realistic, dependency-driven motion trajectories.

Overall, DYNAMO demonstrates robust performance across assemblies with 2 to 7 parts, consistently producing physically plausible and temporally coherent predictions, even in complex or symmetric configurations.