A Bayesian Framework for Active Tactile Object Recognition, Pose Estimation and Shape Transfer Learning

TL;DR

This paper addresses the problem of active tactile object recognition, pose estimation and shape transfer learning, where a customized particle filter and Gaussian process implicit surface is combined in a unified Bayesian framework.

Abstract

As humans can explore and understand the world through active touch, similar capability is desired for robots. In this paper, we address the problem of active tactile object recognition, pose estimation and shape transfer learning, where a customized particle filter (PF) and Gaussian process implicit surface (GPIS) is combined in a unified Bayesian framework. Upon new tactile input, the customized PF updates the joint distribution of the object class and object pose while tracking the novelty of the object. Once a novel object is identified, its shape will be reconstructed using GPIS. By grounding the prior of the GPIS with the maximum-a-posteriori (MAP) estimation from the PF, the knowledge about known shapes can be transferred to learn novel shapes. An exploration procedure based on global shape estimation is proposed to guide active data acquisition and terminate the exploration upon sufficient information. Through experiments in simulation, the proposed framework demonstrated its effectiveness and efficiency in estimating object class and pose for known objects and learning novel shapes. Furthermore, it can recognize previously learned shapes reliably.

Figures (11)

• Figure 1: The proposed framework consists of a customized particle filter and a Gaussian process implicit surface (GPIS). The customized particle filter estimate the joint distribution of object pose and object class upon newly obtained tactile observation, from which the maximum a posteriori (MAP) combination of object class and pose can be extracted. The MAP is used as a prior for GPIS reconstruction when a novel object is identified. An exploration procedure based on the global shape estimation (MAP/GPIS), including target point selection and contact enforcement is proposed to perform active data acquisition. Tactile exploration continues until the termination criterion is met. The learned GPIS from a novel object can be added as a new prior, which enables the framework to recognize it in future exploration.
• Figure 2: An example of sampling particles with point-pair features. For each observed data point pair, $\theta_{i,j},\theta_{i,d},\theta_{j,d}$ and $||\mathbf{x}_{i,j}||_2$ are calculated as the point pair features. $\mathbf{x}_{i,j}$ denotes the vector $\mathbf{x}_j-\mathbf{x}_i$. $\theta_{i,j}$ ,$\theta_{i,d}$, $\theta_{j,d}$ are the angle between normal vectors of $\mathbf{n}_i$ and $\mathbf{n}_j$, the angle between $\mathbf{n}_i$, and $\mathbf{x}_{i,j}$, the angles between $\mathbf{n}_j$ and $\mathbf{x}_{i,j}$ respectively. Point pairs on all known models with similar features are extracted. Finally, by aligning known model point pairs with the observed point pair, possible combinations of the object class and object pose that match the observed point pair are found. $c_s$, $c_e$, $c_r$ represent three object classes, namely square, ellipse, and rectangular respectively. $\mathbf{p_{s}^{0}}$, $\mathbf{p_{e}^{0}}$, $\mathbf{p_{r}^{0}}$ are the original object pose, by convention set to $[0,0,0,0,0,0]$, whereas $\mathbf{p_{s}^{'}}$, $\mathbf{p_{e}^{'}}$, $\mathbf{p_{r}^{'}}$ are the pose after the alignment.
• Figure 3: Example of the contact enforcement procedure. Starting from the target point (red dot), the sensor first moves towards the interior of the MAP shape (purple dot). If no contact is found, a non contact point is recorded and the sensor moves towards the closest existing contact point (green dot). If no contact is established either, the closest known contact point is contacted again. From there, the algorithm takes small steps on the surface along the local tangent plane, towards the target point while remaining in contact with the surface. After a short distance, a new contact point will ultimately be recorded (yellow dot).
• Figure 4: The ten known objects used in the experiments. From left to right, top to bottom, the objects are named as armadillo, asian dragon, elephant, bottle 1, mug 1, ice cream 1, guitar 1, vase 1, office chair and sofa 1 respectively.
• Figure 5: The ten novel objects used in the experiments. From left to right, top to bottom, the objects are neptune, dragon, noisy dino, bottle 2, mug 2, ice cream 2, guitar 2, vase 2, home chair and sofa 2 respectively.