cs.ROMar 7, 2026

A Distributed Gaussian Process Model for Multi-Robot Mapping

Authors: Seth Nabarro, Mark van der Wilk, Andrew J. Davison

Organizations: Dyson Robotics Lab, Imperial College London · Department of Computer Science, University of Oxford

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

We propose DistGP: a multi-robot learning method for collaborative learning of a global function using only local experience and computation. We utilise a sparse Gaussian process (GP) model with a factorisation that mirrors the multi-robot structure of the task, and admits distributed training via Gaussian belief propagation (GBP). Our loopy model outperforms Tree-Structured GPs \cite{bui2014tree} and can be trained online and in settings with dynamic connectivity. We show that such distributed, asynchronous training can reach the same performance as a centralised, batch-trained model, albeit with slower convergence. Last, we compare to DiNNO \cite{yu2022dinno}, a distributed neural network (NN) optimiser, and find DistGP achieves superior accuracy, is more robust to sparse communication and is better able to learn continually.

Figures & tables

Fig. 1 : In DistGP , each robot maintains a block of inducing points which parameterise their local GP (far-left) and into which observations can be fused. Communicating robots harmonise models by connecting inducing points via GP consistency factors and exchanging GBP messages (far-right). Here, the map at each point is predicted by the closest robot. Before communication (centre-left) the map is discontinuous and inaccurate, but this is resolved using only local inter-robot exchanges (centre-right).
Fig. 2 : Limitations of TSGP. Some robots remain unconnected to ensure tree connectivity (as in ). This leads to discontinuities in the prediction at the boundary between unconnected robots ( ), which can be reduced significantly by adding loop-forming edges ( ) (see white rectangles).
Fig. 3 : Benefit of additional edges. The prediction error (vertical axis) reduces as more edges are added to the TSGP model (horizontal axis, x=0x=0 is TSGP), but plateaus after ∼9\sim 9 extra edges, where accuracy is equal to a dense GP prior on all inducing points (green). Results averaged over 1010 seeds, errorbars cover ±1\pm 1 standard error (SE) around mean.
Fig. 4 : Asynchronous + distributed vs batch + centralised model. We compare i) online learning, ad-hoc inter-robot connections, ad-hoc GBP (blue) against ii) a centralised implementation with synchronous GBP sweeps on batch data (red, orange, green). The former converges to the same performance as the centralised, dense prior (green). Averages over 1010 seeds, error bars cover ±1\pm 1 SE around the mean.
RegionDistGP (Ours)DiNNO
Central Atlantic0.134±0.001\bm{0.134\pm 0.001}0.175±0.0010.175\pm 0.001
North Indian Ocean0.160±0.001\bm{0.160\pm 0.001}0.222±0.0010.222\pm 0.001
South Indian Ocean0.230±0.001\bm{0.230\pm 0.001}0.316±0.0020.316\pm 0.002
West Pacific0.108±0.001\bm{0.108\pm 0.001}0.166±0.0010.166\pm 0.001
East Pacific0.240±0.002\bm{0.240\pm 0.002}0.303±0.0010.303\pm 0.001
TABLE I : SST prediction MSE. Error in predicted temperatures over the map, averaged over all timesteps after a 2525 day “burn-in”. Intervals cover ±\pm SE over 44 seeds.
Comm. SettingDistGP (Ours)DiNNO
Default0.134±0.001\bm{0.134\pm 0.001}0.175±0.0010.175\pm 0.001
Comm. Interval 1→51\rightarrow 50.138±0.001\bm{0.138\pm 0.001}0.249±0.0010.249\pm 0.001
Comm. Interval 1→101\rightarrow 100.143±0.001\bm{0.143\pm 0.001}0.294±0.0010.294\pm 0.001
dcommd_{\mathrm{comm}} 1500km→1100km1500\mathrm{km}\rightarrow 1100\mathrm{km}0.143±0.001\bm{0.143\pm 0.001}0.420±0.0050.420\pm 0.005
dcommd_{\mathrm{comm}} 1500km→700km1500\mathrm{km}\rightarrow 700\mathrm{km}0.169±0.001\bm{0.169\pm 0.001}−†-^{\dagger}
TABLE II : Varying communication. MSE with different communication settings for the Central Atlantic region. † optimisation diverged. Intervals cover ±\pm SE over 44 seeds.
Fig. 5 : Distributed occupancy mapping. Our method converges to an accurate map in a single pass through the robots’ trajectories, as observations are fully fused into the model. In contrast, DiNNO trains NNs iteratively and therefore requires many passes to overcome catastrophic forgetting. We note that our map at convergence is also more accurate than DiNNO’s.
Fig. 6 : Test error during mapping (log-scale horizontal axis).

Explore similar work

CardsList

  1. Resource-Efficient Distributed Recursive Gaussian Processes

    Sep 22, 2026Josephine King, Ali Emre Balci, Raj Thilak RajanDistributed OptimizationGaussian Processes

    Signal Processing Systems, Faculty of EEMCS, Delft University of Technology, Mekelweg 4, 2628CD, Delft, The Netherlands

  2. Consensus-based Recursive Multi-Output Gaussian Process

    Apr 11, 2026Yogesh Prasanna Kumar Rao, Tamas Keviczky, Raj Thilak RajanGaussian ProcessesGaussian Process Regression

  3. DD-GBP: Anytime Multi-Arm Planning and Continuous Execution via Distributed Dynamic Gaussian Belief Propagation

    Oct 14, 2026Zhao Jin, Fengyi Dai, Yuetong Yin +3Motion PlanningMulti-Robot Motion Planning

    State Key Laboratory of Tribology in Advanced Equipment, Department of Mechanical Engineering, Tsinghua University, Beijing, 100084, China · Beijing Key Laboratory of Transformative High-end Manufacturing Equipment and Technology, Department of Mechanical Engineering, Tsinghua University, Beijing 100084, China · State Key Laboratory of High-end Heavy-load Robots, Midea Group, Foshan 528300, China