Reduced latent belief spaces for active perception in robotics
- Publication Type:
- Thesis
- Issue Date:
- 2024
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
The development of a robot’s capability to actively select sensing actions that improve environmental understanding is the concern of active perception research. A common approach is to design sensing policies that minimise the uncertainty in a robot's belief about a state. Planning in the space of all possible beliefs rather than physical space is then a prudent approach to active perception, as it places uncertainty minimisation at the heart of planning strategies. However, due to the size of belief spaces, it is generally infeasible to find optimal plans within the real-time requirements of robotics. This thesis proposes reduced latent belief space planning for efficient active perception. Here, a partially observable latent variable is introduced that satisfies two properties: first, that the original state of interest can be inferred from it, and second, that its belief space is simpler than original. Then, planning in this reduced belief space to improve latent variable estimation is an efficient proxy for estimating the original state. To evaluate this theoretical paradigm shift, we present a suite of reduced latent belief space planning algorithms for active perception. Empirical validation and derivation of performance guarantees for each demonstrates our approach is a promising advancement in active perception.
Please use this identifier to cite or link to this item: