[Paper Review] Simultaneous View and Feature Selection for Collaborative Multi-Robot Perception

Brian Reily, Hao Zhang|arXiv (Cornell University)|Dec 17, 2020

Robotics and Sensor-Based Localization4 citations

TL;DR

This paper proposes a unified regularized optimization framework that simultaneously performs view and feature selection in collaborative multi-robot perception, using sparsity-inducing norms to identify the most representative robots and discriminative sensor modalities. The method achieves accurate object recognition and effective robot task allocation through a novel, convergent iterative solver, validated in simulation and on a physical robot system.

ABSTRACT

Collaborative multi-robot perception provides multiple views of an environment, offering varying perspectives to collaboratively understand the environment even when individual robots have poor points of view or when occlusions are caused by obstacles. These multiple observations must be intelligently fused for accurate recognition, and relevant observations need to be selected in order to allow unnecessary robots to continue on to observe other targets. This research problem has not been well studied in the literature yet. In this paper, we propose a novel approach to collaborative multi-robot perception that simultaneously integrates view selection, feature selection, and object recognition into a unified regularized optimization formulation, which uses sparsity-inducing norms to identify the robots with the most representative views and the modalities with the most discriminative features. As our optimization formulation is hard to solve due to the introduced non-smooth norms, we implement a new iterative optimization algorithm, which is guaranteed to converge to the optimal solution. We evaluate our approach through a case-study in simulation and on a physical multi-robot system. Experimental results demonstrate that our approach enables effective collaborative perception through accurate object recognition and effective view and feature selection.

Motivation & Objective

• To address the gap in collaborative multi-robot perception where view and feature selection are treated independently rather than jointly.
• To develop a unified optimization formulation that integrates multi-view object recognition, view selection, and feature selection into a single framework.
• To enable intelligent robot coordination by identifying which robots provide the most informative views and which sensor modalities are most discriminative.
• To design a convergent iterative algorithm for solving the non-smooth, non-convex optimization problem arising from structured sparsity regularization.
• To validate the approach in both high-fidelity simulation and real-world multi-robot experiments for robustness and accuracy.

Proposed method

• The method formulates collaborative perception as a regularized optimization problem with structured sparsity-inducing norms to promote selection of key robots and features.
• It uses l1-norms on robot weights and mixed-norms (group lasso-like) on feature modalities to induce sparsity and select representative views and modalities.
• The optimization combines data fidelity (reconstruction error) with regularization to balance accuracy and sparsity in view and feature selection.
• A novel iterative algorithm is proposed, proven to converge to the optimal solution through successive minimization of a surrogate objective function.
• The algorithm alternates between updating robot weights and feature modalities using proximal operators derived from the regularization terms.
• The method is implemented with a block-coordinate descent approach that ensures monotonic decrease in the objective function at each iteration.

Experimental results

Research questions

• RQ1 Can a unified optimization framework simultaneously select the most informative robot views and the most discriminative sensor features for collaborative object recognition?
• RQ2 How does structured sparsity via regularization improve recognition accuracy while reducing reliance on redundant or occluded observations?
• RQ3 What is the impact of joint view and feature selection on robot task allocation and system efficiency in multi-robot systems?
• RQ4 Can a novel iterative solver with theoretical convergence guarantees be designed for non-smooth, non-convex optimization in this context?
• RQ5 How does the proposed method compare to independent view and feature selection in terms of recognition performance and computational efficiency?

Key findings

• The proposed method achieves high object recognition accuracy by selecting only the most informative robot views and discriminative sensor modalities.
• The iterative solver converges to an optimal solution, as proven theoretically, with monotonic decrease in the objective function at each iteration.
• Experimental results in simulation show that the method outperforms baseline approaches in recognition accuracy while reducing the number of active robots.
• On the physical multi-robot system, the framework successfully identified key robots and modalities, enabling effective perception despite occlusions.
• The use of structured sparsity-inducing norms significantly improved selection performance, reducing reliance on redundant or low-quality observations.
• The method enables dynamic robot reassignment, as robots with less representative views can be redirected to observe new targets.