Gaussian Splatting and Point Cloud-Based Workspace Prediction for Collision-Free Trajectory Planning in Collaborative Robots
Jungho Seo, DongWook Kim
AI summary
Problem
Collaborative robots lack real-time situational awareness and collision avoidance in dynamic or partially occluded industrial environments, while existing vision-language models impose prohibitive computational overhead for real-time deployment.
Approach
The framework reconstructs the workspace using 3D Gaussian Splatting, estimates robot poses with a GICP-FGR pipeline, predicts joint angles via a DGCNN, forecasts future trajectories using Dynamic Mode Decomposition, and enforces safety in real-time with Control Barrier Functions.
Key results
- GICP-FGR pipeline resolves ICP singularity and rotation redundancy for stable 6-DoF pose estimation
- DMD accurately predicts future robot poses from only two or three prior observations
- DGCNN reduces joint angle regression MAE by 45–53% using augmented 3D-GS point cloud data
- CBF-based QP successfully filters control inputs to guarantee collision-free operation
Why it matters
Enables real-time, computationally lightweight collision avoidance and workspace prediction for industrial collaborative robots operating in dynamic or partially occluded environments.
Abstract
As multi-robot collaboration becomes increasingly prevalent in modern industrial settings, ensuring collision-free operation among robots sharing the same workspace remains a critical challenge. This paper proposes an integrated framework that combines 3D Gaussian Splatting (3D-GS) for high-fidelity scene reconstruction, Generalized Iterative Closest Point (GICP) with Fast Global Registration (FGR) for robust pose estimation, a Deep Graph Convolutional Neural Network (DGCNN) for joint angle regression from point cloud data, Dynamic Mode Decomposition (DMD) for trajectory prediction, and a Con- trol Barrier Function (CBF) for real-time safety enforcement. Through experiments, we validated the trajectory prediction of 0- DOF objects and confirmed that joint angle prediction is feasible from 3D-GS-based PLY data using DGCNN-based regression, with training data collected at joint angle intervals of 15◦–45◦.