DOME: Distributed Online Learning based Multi-Estimate Fusion for Cooperative Predictive Target Tracking Using a Robotic Swarm

Shubhankar Gupta, Saksham Sharma, Suresh Sundaram

Abstract: This paper investigates cooperative predictive target tracking using a robotic swarm operating under high prediction bias and communication uncertainty. The robots interact over a randomly time-varying communication network and exhibit heterogeneity in onboard sensors and prediction algorithms. To address these challenges, a Distributed Online learning-based Multi-Estimate (DOME) fusion algorithm is proposed, which performs a collaborative weighted fusion of local and socially shared predictions. The fusion weights are adapted online using feedback from a prediction loss. Theoretical analysis establishes that conditional expectations of the fusion weights converge under reasonable assumptions. Simulation studies demonstrate that DOME outperforms both covariance-based and online learning-based decentralized fusion baselines, achieving $74\%$ and $72.4\%$ lower prediction loss in performance and scalability tests, respectively -- particularly under conditions involving significant model drift and communication unreliability. Further, DOME fusion is implemented in a ROS-Gazebo simulation environment.

Submission Type: Long submission (more than 12 pages of main content)

Changes Since Last Submission:

The following modifications/additions have been made in the revised version of the paper:

1. A new Related Works section just after the Introduction, with relevant text from the previous Introduction shifted to this new section.

2. Justification for the use of baselines as state-of-the-art in practice in the new Introduction section.

3. A conceptual overview subsection in section 4, giving an overview of DOME, and an overview diagram as Fig. 10 in the Appendix. This overview explains the motivation behind the periodic reset mechanism.

4. Included all the missing variables in the Nomenclature (Appendix A.1).

5. Modified the Broader Impact Statement.

6. Added Ablation Study (check section 6, and Appendix A.3.5, Fig. 6).

7. Added Robustness Study (check section 6), i.e., varying communication failure rate (Appendix A.3.6, Fig. 7), prediction drift rate (Appendix A.3.7, Fig. 8), and noisy observations (Appendix A.5, Fig. 9).

8. Added discussion regarding Sensitivity to Mismatch in $T_o$ (section 6, and Appendix A.6).

9. Added detailed discussion regarding Full Observability Assumption’s Validity in Practice (section 2, and Appendix A.4 ).

10. Added a discussion on Assumptions Regarding the Bias Structure in section 6 as Remark 9.

11. Added a discussion regarding Practical Implications of Assuming Access to True Target State as Remark 4 in section 4.

12. Added a detailed theoretical and empirical analysis of the impact of noisy observations on the DOME weight updates in Appendix A.5.

13. Added a new baseline: Separate Bias Kalman Consensus Fusion (SBKCF), inspired by Separate Bias Kalman Filter (SBKF). Turns out that SBKCF is the best among other baselines. Modified the quantitative results in the abstract, introduction, performance evaluation, and conclusion sections of the paper to reflect this. Added a brief qualitative discussion regarding its performance and modified the plots (Fig. 3a and 3b) in the performance evaluation section. Details regarding SBKCF are added in Appendix A.3.2.

14. The main paper has become 14 pages long. Changed the submission type to `long submission' accordingly.

Video: https://youtu.be/-8IKBRZDQVQ

Code: https://github.com/airl-iisc/dome

Assigned Action Editor: ~Sai_Aparna_Aketi1

Submission Number: 6466