Traditional joint-link robots have been widely used in production lines because of their high precision for single tasks. With the development of the manufacturing and service industries, the requirement for the comprehensive performance of robotics is growing. Numerous types of bio-inspired robotics have been investigated to realize human-like motion control and manipulation. A study route from inner mechanisms to external structures is proposed to imitate humans and animals better. With this idea, a brain-inspired intelligent robotic system is constructed that contains visual cognition, decision-making, motion control, and musculoskeletal structures. This paper reviews cutting-edge research in brain-inspired visual cognition, decision-making, motion control, and musculoskeletal systems. Two software systems and a corresponding hardware system are established, aiming at the verification and applications of next-generation brain-inspired musculoskeletal robots.

Organisms have an innate ability to rapidly produce diverse and flexible movement. Biological motor systems are composed of highly redundant muscle actuators and have strongly nonlinear dynamic characteristics. How organisms cope with such complex system is still an open question. Motor primitive theory proposed by neuroscientists has provided a convincing explanation for this extraordinary ability and have supported by several biological evidence. In this paper, based on the gain primitive model of cortical network proposed by neuroscientists, new algorithms for learning and combining gain primitives were designed to control musculoskeletal and robotic system with highly redundant actuators. It can realize dimensionality reduction of controll from the number of actuators to the number of primitives. A parameter adaptation algorithm inspired by monoamines modulation mechanism was applied to improve the learning efficiency of gain primitive. Learned motion experiences and primitives were introduced to effectively generate new movements. Validation experiments were carried out on a musculoskeletal model with 9 muscles and an articulated robot Baxter with redundant actuators. The experimental results demonstrated that the recurrent neural network modulated by gain primitives can generate high-dimensional control signals and realize efficient motion generalization for practical systems. Note to Practitioners —This paper applied a biologically plausible control method with gain primitive to practical systems with highly redundant actuators. To reduce the cumulative error occurring in learning gain primitive for control networks with high dimensional output, a bioinspired Hebbian learning rule was applied to train a recurrent neural network for optimizing the gain primitive. During the training process, systematic entropy was introduced as an index to adaptively adjust learning parameters, so as to reduce the exploration cost and improve the learning stability of the system. To realize easy implementation and generalization, prior motion experiences were introduced into the initialization of the optimization algorithm, and motion generalization was realized by using the particle swarm optimization algorithm. The algorithms designed in this paper are applicable to complex systems with highly redundant actuators such as humanoid robots.

Humans have remarkable physical abilities and neurological coordination, allowing them to perform an extensive variety of movements, most of which are cyclic or quasi-cyclic. Numerous methods have been proposed to achieve cyclic movements in humanoid robots, such as central pattern generators, model predictive control, imitation learning, and reinforcement learning. However, these methods mainly achieve a suboptimal single-gait or a restricted range of gaits, constraining the humanoid’s movement capabilities. To address this issue, we proposed a universal gait template generation method. It allows continuous variation in gait parameters, such as cycle period, swing time ratio, body height, and maximum foot elevation. Through sampling within this space, various gaits can be generated, such as standing, walking, speed walking, running, and jumping. Deep reinforcement learning is employed to track the generated randomized gait template as reference trajectories within a single network. This method allows the humanoid to achieve infinitely many physically feasible gaits and select the optimal one. We conducted simulation analyses and achieved robust control with zero-shot sim-to-real transfer on the Q-series humanoid robots’ hardware platform, achieving a maximum speed of 2 m/s, payload capacity of 20 kg, and a maximum impact resistance of 80 Ns.

Simultaneous localization and mapping (SLAM) is the basis for intelligent robots to explore the world. As a promising method for 3D reconstruction, 3D Gaussian splatting (3DGS) integrated with SLAM systems has shown significant potential. However, due to environmental uncertainties, errors in the tracking process with 3D Gaussians can negatively impact SLAM systems. This paper introduces a novel dense RGB-D SLAM system based on 3DGS that refines Gaussians through sub-Gaussians in the camera coordinate system. Additionally, we propose an algorithm to select keyframes closely related to the current frame, optimizing the scene map and pose of the current keyframe. This approach effectively enhances both the tracking and mapping performance. Experiments on high-quality synthetic scenes (Replica dataset) and low-quality real-world scenes (TUM-RGBD and ScanNet datasets) demonstrate that our system achieves competitive performance in tracking and mapping.