GaussLite: Task-conditioned 3D mapping boosts ROI PSNR by 2.72 dB

GaussLite tackles a core inefficiency in 3D Gaussian Splatting (3DGS) for robotics: existing systems uniformly distribute representation capacity across an entire scene, wasting compute on irrelevant geometry. Developed by Annika Thomas, Mason Peterson, and Jonathan P. How at MIT, GaussLite introduces a task-conditioned approach that dynamically allocates resources based on a natural-language task specification. Given a posed RGB-D stream and a command like "prepare to pick up the object on the desk," a one-shot LLM parser extracts target and anchor objects. These are grounded per-frame by an open-vocabulary detector and segmented to produce real-time per-pixel relevance masks. The mapper then directs seeding density, gradient flow, and scaling toward the most relevant parts of the scene.

The results are compelling: at a matched Gaussian budget and real-time mapping at 4 Hz on resource-constrained hardware, GaussLite outperforms baselines on ROI PSNR by an average of +2.72 dB on the Replica dataset and +2.23 dB in real-world indoor and outdoor demonstrations. The system also enables efficient multi-agent map fusion: two task-specialized agents' maps can be merged via per-voxel voting on active-optimization counts, achieving +3.42 dB improvement over simple concatenation while sharing only 7.08% of the map on average. This work points toward more efficient, task-aware robotic perception that could enable real-time operation on edge devices.