SAFER-Splat (Simultaneous Action Filtering and Environment Reconstruction) is a real-time, scalable, and minimally invasive action filter, based on control barrier functions, for safe robotic navigation in a detailed map constructed at runtime using Gaussian Splatting. We propose a novel Control Barrier Function (CBF) that not only induces safety with respect to all Gaussian primitives in the scene, but when synthesized into a controller, is capable of processing hundred of thousands of Gaussians while maintaining a minimal memory footprint and operating at 15 Hz during online Splat training. To showcase the safety filter, we also introduce SplatBridge, an open-source ROS bridge for real-time GSplat training for robots. We demonstrate the safety and robustness of our pipeline first in simulation, where our method is 20-50x faster, safer, and less conservative than competing methods based on neural radiance fields. We then demonstrate simultaneous GSplat mapping and safety filtering on a drone hardware platform using only on-board perception. We verify that under teleoperation a human pilot cannot invoke a collision.
SAFER-Splat leverages a novel CBF, closely integrated with the GSplat representation, to derive a lightweight, minimally-invasive safe action filter. To synthesize a controller, we embed the CBF safety constraint into a quadratic program, minimizing the deviation between the desired and actuated control effort. By pruning the CBF constraints to identify a minimal number of constraints, our method scales efficiently to run in real-time with hundred of thousands ellipsoidal primitives in the scene. Together, the safety guarantees, computational efficiency, and scalability of SAFER-Splat enable its superior performance in collision-free robot control in online GSplat scenes.
To demonstrate the performance of SAFER-Splat, we introduce SplatBridge. SplatBridge utilizes an incoming stream of RGB images, time-of-flight (ToF) point clouds, and estimated poses from a robot's onboard cameras to train a GSplat in real-time. SplatBridge builds upon NerfBridge, to interface with the GSplat implementation in Nerfstudio. At every keyframe, SplatBridge loads the image into a running buffer and seeds Gaussian primitives from the ToF point-cloud, while asynchronously optimizing the Gaussian attributes and keyframe camera poses. Here, we show the remarkable photorealistic rendering capability of SplatBridge in a real-world scene.
In simulation and hardware experiments, we demonstrate the safety and computational efficiency of SAFER-Splat. Here, we present a few hardware experiment results, where a human pilot repeatedly attempts to force a collision of the drone with obstacles in the scene. We refer readers to our paper for a complete discussion of the experiment results.
Although the human pilot tries to force a collision, the CBF preserves the drone's safety by filtering the control inputs sent to the drone.
SAFER-Splat: Safety with Control Barrier Functions in Online Gaussian Splatting Maps
@misc{chen2024safer-splat,
title={SAFER-Splat: Safety with Control Barrier Functions in Online Gaussian Splatting Maps},
author={Timothy Chen and Aiden Swann and Javier Yu and Ola Shorinwa and Riku Murai and Monroe Kennedy III and Mac Schwager},
journal={arXiv preprint arXiv:2409.09868},
year={2024}
}