Autonomous navigation, multi-target tracking and data registration are estimation problems relevant to a wide range of applications. For robotic navigation, feature-based simultaneous localization and mapping (SLAM) is critical, in which a varying number of features/landmarks is detected for integration into a dynamic map. In 3D reconstruction and scene analysis, point cloud data (PCD) registration is a fundamental necessity and in space situational awareness (SSA), multi-target tracking is of major importance, since the space debris population is so high, that chain reaction collisions, known as the Kessler syndrome, are increasingly likely.
  Although these applications, as well as their operating environments, are highly distinct, they can be addressed with a common underlying random finite set (RFS)-based, multi-object state estimation framework. At ground level and in space, multi-object state estimation is challenging because of 1. Unknown Numbers: The total number of objects fluctuates since ground-based features or space debris appear (birth/spawning) and disappear (death/occlusion) over time; 2. Noisy sensors: Raw sensor measurements contain spatial errors, false alarms, and missed detections and 3. Data association: State-of-the-art solutions must determine which measurement belongs to which specific object, for successful navigation and multi-target tracking.
  Standard solutions to these problems are based on the use of random vectors (RVs) in stochastic filtering, smoothing, or optimization-based approaches. However, standard approaches are often unable to cope with a population of objects whose size is constantly changing. Therefore, recent solutions to these problems have been formulated based on Bayesian RFSs, which model the multi-object configuration itself as a single random variable. Contrary to their RV counterparts, Bayesian RFS methods adopt finite set statistics (FISST), capable of estimating target/map feature number, as well as the state values themselves. They also circumvent the necessity of, often fragile, measurement-to-state data association decisions and target/map feature initiation and termination routines.
  This presentation will focus on new multi object-based solutions to SLAM, 3D-PCD and SSA. It will be shown that the SLAM state can be modelled as a mixed distribution, which jointly represents the vector-valued trajectory and the RFS-valued map, yielding competitive, and often superior results while circumventing the need for fragile data association methods. The 3D PCD registration problem will be recast as a multi-object metric minimization problem, which will be shown to provide robust and accurate solutions under low overlap and high occlusion scenarios. Finally, FISST based SSA in a low Earth orbit environment, in which multiple space debris are detected via telescopic measurements and tracked, utilizing Riemannian manifold geometry will be demonstrated.

This plenary reviews our team’s progression from collective swarm estimation to individual tracking and coordinated sensing. We first develop a single-object Bayesian filter whose composite state includes dynamic and formation variables. Sequential Monte Carlo (SMC) prediction propagates particles through motion and formation models, while geometric likelihoods update their weights. A jump-Markov model handles transitions between different formations. Next, we couple the swarm estimator with a labelled multi-Bernoulli (LMB) random finite set filter. Each member is represented by a labeled Bernoulli component carrying an existence probability and a conditional state density. Geometric compatibility factors incorporate estimated swarm information into the single-object likelihood, improving measurement association, cardinality estimation and identity continuity despite dense spacing, clutter and missed detections. We then develop a mixed representation combining an SMC swarm density with an LMB density for separated objects. Predicted formation geometry and collective motion determine adaptive birth densities near the swarm boundary. Geometry-informed measurement gating selects returns appropriate for departure tracking. We also consider extensions for objects that intermittently join or rejoin the swarm. Finally, we formulate multi-sensor control as a partially observable Markov decision process. A one-step belief policy maximizes a reward balancing swarm-boundary detectability, existence-weighted individual-object detectability and sensor actuation costs. Normalizing the individual-object term by expected cardinality makes it comparable with collective coverage. A central node coordinates prediction and sensor actions, while sensors perform local updates. Generalized covariance intersection combines local posterior densities, with adaptive fusion weights reflecting sensor informativeness.

We have developed various live support devices targeting patients, frail individuals, the elderly, and healthy workers, based on the concept of inclusive design. If development is carried out solely by engineers who do not understand the field, there is a risk of creating something that is not actually usable. Therefore, from the outset of development, it is necessary to bring together members who are familiar with the actual use cases, such as patients, their families, and medical staff. We have already commercialized devices for gait training for hemiplegic patients and lifting assistance suits for industrial workers, and we will introduce these and many other devices under development.