A robot's instantaneous sensory observations do not always reveal task-relevant state information. Under such partial observability, optimal behavior typically involves explicitly acting to gain the missing information. Today's standard robot learning techniques struggle to produce such active perception behaviors. We propose a simple real-world robot learning recipe to efficiently train active perception policies. Our approach, asymmetric advantage weighted regression (AAWR), exploits access to "privileged" extra sensors at training time. The privileged sensors enable training high-quality privileged value functions that aid in estimating the advantage of the target policy. Bootstrapping from a small number of potentially suboptimal demonstrations and an easy-to-obtain coarse policy initialization, AAWR quickly acquires active perception behaviors and boosts task performance. In evaluations on 8 manipulation tasks on 3 robots spanning varying degrees of partial observability, AAWR synthesizes reliable active perception behaviors that outperform all prior approaches. When initialized with a "generalist" robot policy that struggles with active perception tasks, AAWR efficiently generates information-gathering behaviors that allow it to operate under severe partial observability for manipulation tasks. Website: https://penn-pal-lab.github.io/aawr/

Aiming at the three-dimensional (3-D) imaging task under the planar synthetic aperture system, inspired by the classical Range Doppler Algorithm, this paper proposes a novel 3-D imaging algorithm based on the wavenumber domain correction, which has a simpler and more intuitive imaging process and faster computational efficiency. This paper analyzes the expression of 3-D echo data in the horizontal frequency (wavenumber) domain, vertical frequency domain, and range domain. The result indicates that the range migration of the target in the 3-D frequency domain is in the shape of an arc surface, and the range migration equation has a very simple mathematical form. The signal can be easily decoupled in three dimensions by range migration correction, and then the two-dimensional (2-D) inverse Fourier transform is used to reconstruct the 3-D image. The effectiveness of the proposed algorithm is verified by simulation and measured data.

Three-dimensional point cloud registration is a crucial technique for point cloud processing. Iterative closest point (ICP) is widely used for rigid registration of point clouds because of its simplicity but suffers from slow convergence and the tendency to fall into local optimization. In this article, we proposed a new robust and fast point cloud registration method, the ICP optimized by geometric features (including normal, curvature, and point distance), called GFOICP. GFOICP statistically selects registration points by the cross entropy of geometric features of the points, then matches correspondences based on a variable distance threshold, and filters out correct correspondences using an iterative strict constraint on geometric feature similarity. In addition, geometric feature similarity is added as a constraint to the objective function to ensure strict convergence. GFOICP completes registration by iterating correspondence matching, alignment, and convergence judgment. Extensive experiments on publicly available synthetic and real-world datasets have demonstrated that GFOICP significantly improves accuracy and speed compared to standard ICP. GFOICP has similar or higher accuracy and speed than other state-of-the-art registration methods.