Soft robotic operation is a significant trend in daily life and industrial fields. However, soft manipulators are prone to failure during long-term operation under complex working conditions. This paper proposes a novel scheme merging with knowledge intelligent and fault-tolerant control into an alterable grasp configuration soft gripper (AGCSG). Firstly, a knowledge intelligent fault-tolerant soft robotic operation (KIFTSRO) framework with failure detection, knowledge reasoning, and fault-tolerant control is constructed. Next, the available gripper combinations are inferred from all the grasp configurations and different failure cases, and a progressive failure detection strategy is designed to assert the failure cases of sensing modules, and actuation modules as sensing modules are intact or failed. Then, a fault-tolerant soft robotic operation knowledge graph (FTSROKG) with dual input layers is designed to achieve available gripper combination reasoning with ACNet-ConvE algorithm. Finally, the UR5 robot guides the AGCSG to perform complex tasks under the fault-tolerant control scheme. The experimental results reveal that failure detection can be achieved regardless of whether the sensing modules are available or have failed. Moreover, the AGCSG can perform fault-tolerant grasping for different objects with different shapes and hardness, and accomplishes complex tasks including mobile phone charging, tea brewing, kitchen items organizing, daily objects sorting, and items arranging.

Reinforcement learning (RL) has shown promising performance in tackling robotic manipulation tasks (RMTs), which require learning a prolonged sequence of manipulation actions to control robots efficiently. However, most RL algorithms often suffer from two problems when solving RMTs: inefficient exploration due to the extremely large action space and catastrophic forgetting due to the poor sampling efficiency. To alleviate these problems, this paper introduces an Evolutionary Reinforcement Learning algorithm with parameterized Action Primitives, called ERLAP, which combines the advantages of an evolutionary algorithm (EA) and hierarchical RL (HRL) to solve diverse RMTs. A library of heterogeneous action primitives is constructed in HRL to enhance the exploration efficiency of robots and dual populations with new evolutionary operators are run in EA to optimize these primitive sequences, which can diversify the distribution of replay buffer and avoid catastrophic forgetting. The experiments show that ERLAP outperforms four state-of-the-art RL algorithms in simulated RMTs with dense rewards and can effectively avoid catastrophic forgetting in a set of more challenging simulated RMTs with sparse rewards.

Existing soft hand mostly focus on mechanism and material innovation, as well as biological behavior imitation, which have problems of fewer sensing abilities and unfriendly human-robot interaction (HRI). In this paper, we propose a novel bionic soft hand (BSH) with dexterity operation, multi-modal perception and bidirectional HRI. For dexterity operation, we first designed and fabricated a dual-joint bellows (DJB) soft finger, and then, five fingers were integrated into BSH with inertial measurement units (IMUs), contact pressure sensors, curvature sensors and air pressure sensors distributed appropriately. For multi-modal perception, the control and perception system were established to drive BSH in pneumatic mode and process the multi-modal sensor information to perform robust perception capabilities using support vector machine (SVM), extreme learning machine (ELM) and BP neural network (BPNN). For bidirectional HRI, a data glove with force reproduction was utilized to perform posture mapping and tactile force mapping. The experimental results verify that the proposed BSH has a wide workspace range and dexterous operation ability, and can accurately recognize human posture and object properties. Also, hand posture mapping and tactile force mapping have been demonstratedNote to Practitioners—This work was motivated by the research status of existing soft hands that lack sensing abilities and friendly HRI, and aimed to develop a bionic soft hand (BSH) with dexterity operation, multi-modal perception and tactile force interaction. We increased the degrees of freedom of BSH by adopting a bellows structure to build a two-joint soft finger, enhanced the perception ability of BSH by utilizing multi-modal sensors and multiple machine learning algorithms, performed posture mapping and tactile force mapping for bidirectional HRI by designing a novel data glove with multi-modal sensors and a variable stiffness jamming layer. Experimental results verify that the proposed BSH has a wide workspace range and dexterous operation ability, and can accurately recognize human posture and object properties by multi-modal sensors. Future work will focus on imitating the joint structure of human hand and make BSH be closer to the dexterity and perception performance of human hand. Also introducing humanbrain-machine interface is a meaningful challenge.

The utilization of robots in computer, communication, and consumer electronics (3C) assembly has the potential to significantly reduce labor costs and enhance assembly efficiency. However, many typical scenarios in 3C assembly, such as the assembly of flexible printed circuits (FPCs), involve complex manipulations with long-horizon steps and high-precision requirements that cannot be effectively accomplished through manual programming or conventional skill-learning methods. To address this challenge, this article proposes a learning-based framework for the acquisition of complex 3C assembly skills assisted by a multimodal digital-twin environment. First, we construct a fully equivalent digital-twin environment based on the real-world counterpart, equipped with visual, tactile force, and proprioception information, and then collect multimodal demonstration data using virtual reality (VR) devices. Next, we construct a skill knowledge base through multimodal skill parsing of demonstration data, resulting in primitive policy sequences for achieving 3C assembly tasks. Finally, we train primitive policies via a combination of curriculum learning, residual reinforcement learning, and domain randomization methods and transfer the learned skill from the digital-twin environment to the real-world environment. The experiments are conducted to verify the effectiveness of our proposed method.

In this article, an evolution-guided adaptive dynamic programming (EGADP) algorithm is developed to address the optimal regulation problems for the nonlinear systems. In the traditional adaptive dynamic programming algorithms, policy improvement is typically reliant on the gradient information, according to the first order necessity condition. However, these methods encounter limitations when calculating the gradient information becomes infeasible or system dynamics is not differentiable. In response to this challenge, the evolutionary computation is harnessed by EGADP to search for a superior policy during policy improvement. Therefore, compared with the traditional methods, scenarios that gradient information is unavailable can effectively be handled by EGADP. Additionally, the convergence of the algorithm is proven to enhance the rigorousness of the developed method. Finally, the three simulation experiments with realistic physical backgrounds are conducted to comprehensively demonstrate the effectiveness of the established method from different perspectives.