Communication-efficient collaborative 4D mapping via federated temporal Gaussian pruning and knowledge-distilled neural radiance fields for post-disaster rescue robotics

Abstract

Post-disaster rescue operations demand real-time, collaborative 4D mapping of dynamic ruins, yet severely damaged communication infrastructure and limited edge computing resources make conventional neural radiance fields and 3D Gaussian splatting impractical: they lack temporal representation, incur high transmission latency, and do not address the risks of centralizing raw sensor observations. To bridge this gap, we propose Fed-TGP-KD-NeRF, a federated collaborative mapping framework that integrates a learnable temporal latent code with multi-dimensional positional hashing and Fourier feature mapping to construct a continuous 4D dynamic Gaussian field. Adaptive pruning based on Gaussian opacity and gradient sensitivity removes redundant primitives, while a teacher-student knowledge distillation paradigm with intermediate-layer feature alignment and KL divergence compresses the model to approximately 10% of its original size. An improved federated averaging mechanism with dynamic weight adaptation (bandwidth and battery status) and periodic synchronization enables data-local collaboration by keeping raw observations on the originating robot. Experiments on the 3DAeroRelief dataset show that the proposed method achieves a PSNR of 32.58 dB, SSIM of 0.912, and real-time rendering at 35.26 FPS, while reducing per-round communication overhead to only 13.45 MB and converging in 42 rounds, outperforming the evaluated baseline methods. The framework balances reconstruction fidelity, communication efficiency, and raw-data locality; it does not claim a formal differential-privacy guarantee.