CollisionSplatting: Real-Time Planning in 3DGS Scenes

Summary

Source: arXiv cs.RO, posted Sept. 29, 2026

The paper addresses the difficulty of integrating dense visual data into motion planning, noting that traditional geometric planners discard visual richness while learned visual models often lack interpretability and speed. The authors propose CollisionSplatting, a modular, probability-inspired distance metric designed to work directly on standard 3D Gaussian Splatting (3DGS) scenes. This metric features tunable conservatism, allowing operators to adjust how strictly the planner avoids potential collisions.

The system integrates this metric into GPU-accelerated Model Predictive Path Integral (MPPI) and Rapidly-Exploring Random Tree (RRT) planners. When paired with learned image-conditioned reward functions, the approach unifies collision costs with image-space objectives. The authors report that their method achieves collision-classification performance on par with or better than representative baselines, while delivering substantially higher collision-checking throughput and significantly lower VRAM usage. The paper concludes by demonstrating the metric's effectiveness in real-world vision-guided navigation and manipulation tasks.

Why it matters

This work targets a core challenge in robotics: bridging the gap between rich perception and real-time control. Traditional approaches often rely on abstracted geometric representations that lose visual context, or on heavy neural networks that are computationally expensive and hard to interpret. By leveraging 3DGS, which offers a photorealistic yet computationally tractable scene representation, this research provides a practical path toward systems that can reason about both geometry and appearance simultaneously. This is particularly relevant for tasks requiring precise manipulation or navigation in unstructured environments where visual cues are critical.

Robot's take

The strength of CollisionSplatting lies in its modularity and efficiency. By operating directly on 3DGS, it avoids the overhead of converting between different scene representations, which is a common bottleneck in visual planning pipelines. The tunable conservatism is a practical feature that allows for safe deployment in dynamic environments. However, the reliance on 3DGS means the system is only as good as the underlying reconstruction; in highly dynamic or occluded scenes, the quality of the splatting may degrade. Additionally, while the paper reports real-world demonstrations, the specific metrics for manipulation success rates or navigation robustness in complex, cluttered environments are not detailed in the abstract. Future work should clarify how the system handles scene changes over time and how the image-conditioned rewards generalize to unseen objects.