Implicit Articulated Robot Morphology Modeling with Configuration Space Neural Signed Distance Functions

TL;DR

The paper introduces RNDF, a configuration-space neural signed distance function that implicitly encodes articulated robot morphology by conditioning on forward kinematics. RNDF achieves millimeter-level accuracy with substantially fewer parameters and enables differentiable joint-space derivatives for rapid planning, demonstrated on a 7-DOF arm and an arm-hand grasping framework. The approach outperforms baselines in distance prediction and collision-avoidance classification while delivering highly efficient query times, making it suitable as a surrogate model for 3D spatial planning and optimization-based grasping in cluttered environments. Applications to arm-hand grasp planning illustrate unified, constraint-aware planning that integrates contact, reachability, and collision avoidance, with potential for extension to dynamics and noisy sensing.

Abstract

In this paper, we introduce a novel approach to implicitly encode precise robot morphology using forward kinematics based on a configuration space signed distance function. Our proposed Robot Neural Distance Function (RNDF) optimizes the balance between computational efficiency and accuracy for signed distance queries conditioned on the robot's configuration for each link. Compared to the baseline method, the proposed approach achieves an 81.1% reduction in distance error while utilizing only 47.6% of model parameters. Its parallelizable and differentiable nature provides direct access to joint-space derivatives, enabling a seamless connection between robot planning in Cartesian task space and configuration space. These features make RNDF an ideal surrogate model for general robot optimization and learning in 3D spatial planning tasks. Specifically, we apply RNDF to robotic arm-hand modeling and demonstrate its potential as a core platform for whole-arm, collision-free grasp planning in cluttered environments. The code and model are available at https://github.com/robotic-manipulation/RNDF.

Figures (6)

• Figure 1: Visualization of the isosurface at different distance values of the proposed method, which showcases its precision and smoothness in 3D space. The inherent features of the proposed method benefit downstream tasks, and an arm-hand grasp planning framework is proposed.
• Figure 2: The detailed neural network structure of the proposed RNDF for KUKA iiwa 7, which constructs a mapping from $\mathbb{R}^{7+3} \rightarrow \mathbb{R}^{8}$. By incorporating the multi-head and hierarchical feature design, we achieve high-precision distance prediction with a significantly smaller model size.
• Figure 3: The data distribution is balanced through a separate inside and outside sampling design, ensuring accurate boundary modeling.
• Figure 4: Examples of the visualized implicit distance function with random joint configurations built by different methods. The solid and transparent isosurfaces represent the prediction value $\text{min}(f(\mathbf{q},\mathbf{p}))=0.001$ and $\text{min}(f(\mathbf{q},\mathbf{p}))=0.1$ respectively. Our proposed RNDF (with different feature sizes) shows superiority in smoothness, stability, and precision.
• Figure 5: Examples of generated arm-hand pinch and power grasp configuration with different environmental settings. The red sphere, cube, and cylinder represent obstacles that the arm-hand system needs to avoid. The target object is a banana for pinch grasp and an apple for power grasp.