Fast Object Segmentation Learning with Kernel-based Methods for Robotics

Abstract

Multi-fingered robotic hands could enable robots to perform sophisticated manipulation tasks. However, teaching a robot to grasp objects with an anthropomorphic hand is an arduous problem due to the high dimensionality of state and action spaces. Deep Reinforcement Learning (DRL) offers techniques to design control policies for this kind of problems without explicit environment or hand modeling. However, training these policies with state-of-the-art model-free algorithms is greatly challenging for multi-fingered hands. The main problem is that an efficient exploration of the environment is not possible for such high-dimensional problems, thus causing issues in the initial phases of policy optimization. One possibility to address this is to rely on off-line task demonstrations. However, oftentimes this is incredibly demanding in terms of time and computational resources.In this work, we overcome these requirements and propose the A Grasp Pose is All You Need (G-PAYN) method for the anthropomorphic hand of the iCub humanoid. We develop an approach to automatically collect task demonstrations to initialize the training of the policy. The proposed grasping pipeline starts from a grasp pose generated by an external algorithm, used to initiate the movement. Then a control policy (previously trained with the proposed G-PAYN) is used to reach and grab the object. We deployed the iCub into the MuJoCo simulator and use it to test our approach with objects from the YCB-Video dataset. The results show that G-PAYN outperforms current DRL techniques in the considered setting, in terms of success rate and execution time with respect to the baselines. The code to reproduce the experiments will be released upon acceptance.

Learn Fast, Segment Well: Fast Object Segmentation Learning on the iCub Robot

“…In this section, we investigate the impact of region proposal adaptation on the overall performance. In particular, in Section VI-A, we show that, with respect to our previous work [1],…”

“…In addition, we provide an extensive experimental analysis to investigate the training time/accuracy tradeoff on two public datasets (i.e., YCB-Video [3] and HO-3D [4]). In particular, we show that our method is much more accurate than [1], while requiring a comparable training time. Moreover, the proposed method allows to obtain accuracy similar to conventional fine-tuning approaches, while being trained much faster.…”

“…The backbone of the network is trained offline, while the Kernel-based classifiers are adapted online. In this article, we build on our previous work [1], in that we include the adaptation of the region proposal network and a novel training protocol that allows to further reduce the training time. This makes the pipeline suitable for online implementation.…”

“…3. Among these tensors, we consider as positive samples for the n th classifier the features associated to the pixels in the ground-truth masks of class n. Instead, we consider as negative samples the features associated to the background pixels contained in ground-truth bounding boxes of class n. Given the great amount of training samples, to speed-up the training procedure, we randomly subsample both the positive and the negative features by a factor r. According to the analysis provided in [1], we set r to 0.3.…”

“…Specifically, we propose an instance segmentation pipeline that extends and improves our previous work [1]. In [1], we proposed a fast learning method for instance segmentation of novel objects. One limitation of that method was to rely on a pre-trained region proposal network.…”