Robots whose shape and stiffness are determined by internal forces generally have complex shape-stiffness relationships that depend on their structure. As a result, there are difficulties such as a decrease in shape reproducibility when the robot is not stiff, and a decrease in the range of motion when the robot is stiff. In this study, we propose a motion planning method that balances shape and stiffness by learning forward and inverse kinematics using a stochastic neural network (NN) and using the uncertainty that can be evaluated by the NN. Through experiments using a tensegrity manipulator with 40 actuators and 20 degrees of freedom in bending posture, we verify the validity of the proposed method.

While imitation learning provides a simple and effective framework for policy learning, acquiring consistent action during robot execution remains a challenging task. Existing approaches primarily focus on either modifying the action representation at data curation stage or altering the model itself, both of which do not fully address the scalability of consistent action generation. To overcome this limitation, we introduce the Diff-Control policy, which utilizes a diffusion-based model to learn action representation from a state-space modeling viewpoint. We demonstrate that diffusion-based policies can acquire statefulness through a Bayesian formulation facilitated by ControlNet, leading to improved robustness and success rates. Our experimental results demonstrate the significance of incorporating action statefulness in policy learning, where Diff-Control shows improved performance across various tasks. Specifically, Diff-Control achieves an average success rate of 72% and 84% on stateful and dynamic tasks, respectively. Notably, Diff-Control also shows consistent performance in the presence of perturbations, outperforming other state-of-the-art methods that falter under similar conditions. Project page: https://diff-control.github.io/

In flexible redundantly-driven multi-DOF systems, like living beings, the representation of redundant kinematics including the diversity of solutions, is crucial for leveraging its distinctive characteristics. This paper proposes an active learning framework for forward and inverse modeling of complex kinematics that improves expressions of control space, task space, and null space. It consists of a Variational Auto Encoder (VAE)-type network that internally holds expressions of control space, task space, and null space, and an algorithm for selecting new data using the cross-entropy method. The validity of the proposed system was verified using a tensegrity manipulator driven by 40 pneumatic cylinders. As a result, it was confirmed that active learning contributed to achieving the entire range of motion covered and a well-organized representation of the null space.

This paper introduces a novel approach for modeling the dynamics of soft robots, utilizing a differentiable filter architecture. The proposed approach enables end-to-end training to learn system dynamics, noise characteristics, and temporal behavior of the robot. A novel spatio-temporal embedding process is discussed to handle observations with varying sensor placements and sampling frequencies. The efficacy of this approach is demonstrated on a tensegrity robot arm by learning end-effector dynamics from demonstrations with complex bending motions. The model is proven to be robust against missing modalities, diverse sensor placement, and varying sampling rates. Additionally, the proposed framework is shown to identify physical interactions with humans during motion. The utilization of a differentiable filter presents a novel solution to the difficulties of modeling soft robot dynamics. Our approach shows substantial improvement in accuracy compared to state-of-the-art filtering methods, with at least a 24% reduction in mean absolute error (MAE) observed. Furthermore, the predicted end-effector positions show an average MAE of 25.77mm from the ground truth, highlighting the advantage of our approach. The code is available at https://github.com/ir-lab/soft_robot_DEnKF.

This paper uses a data-driven approach to model a highly redundantly driven tensegrity manipulator's forward and inverse kinematics. The tensegrity manipulator is based on a class-1 tensegrity with 20 struts and bends by 40 pneumatic actuators whose internal pressures are independently controlled. Based on the data obtained through random trials with the robot, a VAE-based kinematics model is trained. The forward model, inverse model, and null space of kinematics are simultaneously acquired as subnetworks of the VAE-based kinematics model. Experiments confirmed that the subnetworks representing forward and inverse kinematics could be used for the end position estimation and control, respectively. In addition, the subnetwork representing null space can generate different target pressures that achieve the same end position, which was confirmed to mean variable stiffness properties similar to musculoskeletal robots.

This paper reports a tensegrity manipulator driven by 40 pneumatic cylinders without mechanical springs. In general, tensegrity robots use mechanical springs to achieve a stable/curved tensegrity structure, and this is true even when a component extends/retracts with an actuator. The stiffness of the mechanical spring should be high to increase the stiffness of the entire structure and improve the control response, but low to deform the structure. This fact means that the introduction of mechanical springs causes serious trade-offs in its design and control. In this study, we use pneumatic actuators not only for active deformation but also for passive. In this paper, we introduce the design and control system and then show the difference in response characteristics between the case with and without a spring, demonstrating the importance of the approach without a mechanical spring.

Musculoskeletal robots that are based on pneumatic actuation have a variety of properties, such as compliance and back-drivability, that render them particularly appealing for human-robot collaboration. However, programming interactive and responsive behaviors for such systems is extremely challenging due to the nonlinearity and uncertainty inherent to their control. In this paper, we propose an approach for learning Bayesian Interaction Primitives for musculoskeletal robots given a limited set of example demonstrations. We show that this approach is capable of real-time state estimation and response generation for interaction with a robot for which no analytical model exists. Human-robot interaction experiments on a 'handshake' task show that the approach generalizes to new positions, interaction partners, and movement velocities.

Motor babbling and goal babbling has been used for sensorimotor learning of highly redundant systems in soft robotics. Recent works in goal babbling have demonstrated successful learning of inverse kinematics (IK) on such systems, and suggest that babbling in the goal space better resolves motor redundancy by learning as few yet efficient sensorimotor mappings as possible. However, for musculoskeletal robot systems, motor redundancy can provide useful information to explain muscle activation patterns, thus the term motor abundance. In this work, we introduce some simple heuristics to empirically define the unknown goal space, and learn the IK of a 10 DoF musculoskeletal robot arm using directed goal babbling. We then further propose local online motor babbling guided by Covariance Matrix Adaptation Evolution Strategy (CMA-ES), which bootstraps on the goal babbling samples for initialization, such that motor abundance can be queried online for any static goal. Our approach leverages the resolving of redundancies and the efficient guided exploration of motor abundance in two stages of learning, allowing both kinematic accuracy and motor variability at the queried goal. The result shows that local online motor babbling guided by CMA-ES can efficiently explore motor abundance at queried goal positions on a musculoskeletal robot system and gives useful insights in terms of muscle stiffness and synergy.

It has been widely considered that a distinctive feature of musculoskeletal structures is that both the joint angle and stiffness can be changed by exploiting the agonistantagonist driving of the joint. However, musculoskeletal systems in animals and humans are typically highly complex, and the simple agonist-antagonist driving is rarely found. Therefore, in accordance with the increasing complexity of musculoskeletal robots, the feature that causes the robot to assume a posture with different stiffness values becomes difficult to achieve, owing to the difficulty in modeling the kinematics. Although datadriven approaches such as the neural network are regarded as suitable for modeling complex relationships, the training data are difficult to obtain because measuring joint stiffness is typically extremely difficult in contrast to measuring an actuator's state and posture. Hence, we herein propose the common dimensional autoencoder where the encoded feature exhibits identical dimensions to the original input vector. In the proposed network, in parallel with the original unsupervised training using the data of the actuators' states, supervised training at part of the encoded features is performed using posture data. Consequently, features expressing the redundancy of inverse kinematics appear at the remaining part of the encoded features without using data such as joint stiffness. The validity of the proposed method was confirmed successfully through an experiment using a 10 degrees-of-freedom complex musculoskeletal robot arm driven by pneumatic artificial muscles.

A number of researches have been focusing on the development and control of robots with soft structures such as flexible musculoskeletal systems. Thus far, it has been reported that these robots can achieve high adaptability to environments despite their extremely simple controllers. However, because these robots are difficult to model mathematically, there is still no systematic design policy, in which control theory has been playing a role in conventional robotics, for constituting simple controllers. To tackle this problem, we propose a new approach using a neural network to obtain mathematical models. In particular, with this method, the control theory is applied to linear system models extracted from a network trained to express the forward dynamics of a robot. Through simulations, the validity and advantage of the proposed method was successfully confirmed.

Stochastic resonance (SR) is a phenomenon occurring in some nonlinear systems by which a signal provided as input that is too small in magnitude to normally influence the system's output can actually influence the system's output once a non-zero level of noise is provided. SR has been extensively studied both theoretically and experimentally, and noise has been exploited for improving the performance in both biological and artificial systems. In addition to its scientific importance, the use of noise has attracted interest because of its potential to overcome the present limitations in engineering applications. In this study, we investigate the use of a universal approximator exploiting SR as a new realization of well-established feed-forward neural networks. The proposed universal approximator consists of groups of threshold elements. Although the approximation universality of a network consisting of threshold elements has been proven in terms of extreme learning machine implementations, once SR is taken into account, the system can be modeled in a form identical to that of a classic generic three-layered neural network, for which the universal approximation capability has been proven. The capability of the proposed approximator for serving as a universal approximator is first proven theoretically in the limit of an infinite number of hidden units. Subsequently, the performance achieved by the backpropagation type and the extreme learning machine type learning algorithms is experimentally evaluated for cases involving limited numbers of hidden units, highlighting the SR effect occurring in the proposed system.

The aim of this research is to develop a musculoskeletal robot arm, which has a similar mechanical structure to that of a human, to simply make the robot move similarly to a human based on his/her electromyographic signals. In recent years, many musculoskeletal robots have been developed to show advantages of their bio-inspired designs. In this research, we propose a new perspective of their advantages that the control using biological signals can be simplified thanks to the similarity of mechanical structure. To this end, the shoulder complex, which consists of several bones and joints, is focused because it is the most difficult part to develop in humanlike musculoskeletal robot arms. In particular, we develop a linkage mechanism, which can realize similar function to the shoulder complex, to make a surface electromyography(sEMG) based posture control applicable for this complex system. The advantage, that a simple mapping is still available to control the posture to be the same to that of a human, has been successfully shown in an experiment.

Humans can walk stably and adaptively in the presence of various environmental changes. Their bodies have very complex musculoskeletal structures that contribute to their walking stability and adaptability. In this paper, we focus on the gluteus medius, one of muscles contributing to the support of the pelvis. The gluteus medius supports the pelvis when the leg is in a standing position, but it should not inhibit its smooth swing motion. This mechanism is assumed to be realized by the musculoskeletal structure. This paper is devoted to explaining the mechanism by reproducing it using artificial pneumatic muscles. We developed a musculoskeletal robot with a gluteus medius, and showed that the mechanism is well-designed for realizing two seemingly contradictory functions: being stiff to support the pelvis while standing and being smooth to allow swinging. This finding can be utilized for adaptive walking of musculoskeletal humanoid robots.

This paper proposes a learning from demonstration system based on a motion feature, called phase transfer sequence. The system aims to synthesize the knowledge on humanoid whole body motions learned during teacher-supported interactions, and apply this knowledge during different physical interactions between a robot and its surroundings. The phase transfer sequence represents the temporal order of the changing points in multiple time sequences. It encodes the dynamical aspects of the sequences so as to absorb the gaps in timing and amplitude derived from interaction changes. The phase transfer sequence was evaluated in reinforcement learning of sitting-up and walking motions conducted by a real humanoid robot and compatible simulator. In both tasks, the robotic motions were less dependent on physical interactions when learned by the proposed feature than by conventional similarity measurements. Phase transfer sequence also enhanced the convergence speed of motion learning. Our proposed feature is original primarily because it absorbs the gaps caused by changes of the originally acquired physical interactions, thereby enhancing the learning speed in subsequent interactions.

Direct teaching is suitable for generating motions of robot arms which have complex kinematics. So far, we have proposed a direct teaching method specialized for musculoskeletal robot arms actuated by pneumatic artificial muscles (PAMs) based on their pressure and tension information. In the method, it is important to prevent slacks and excessive tensions of PAMs to efficiently obtain the pressure and the tension information during the teaching phase. In this research, we propose a method for generating active behavior of musculoskeletal robots driven by PAMs to effectively receive human's direct teaching. The method, which is naturally derived from the simple analytical model of PAMs, requires only pressure and tension information of PAMs in musculoskeletal robot arms and does not need to cope with complex inverse kinematics problem. The validity of the method has been confirmed in the experiment using a minimalistic 2DOFs anthropomorphic musculoskeletal robot arm actuated by three pairs of agonist/antagonist PAMs.

Recently, the tendon-driven mechanism with variable joint stiffness has received attention for use in the development of a humanoid robot operated in an uncertain environment with physical contact. In this paper, we propose a mechanism to control the position and joint stiffness of a tendon-driven manipulator independently, using dedicated actuators. This mechanism consists of two parts: a component that transforms the movements of the tendons to activate the actuators, and a component that applies tensile forces to adjust the joint stiffness. We named this mechanism “tendon routing resolving inverse kinematics” (TRIK). The methodology for designing this mechanism for various tendon-driven manipulators is presented with several examples. We designed TRIK for a manipulator with one degree of freedom and nonconstant-moment arms. Finally, experiments of variable joint stiffness with nonlinearly elastic components were conducted to validate the proposed mechanism.

Sarcomere is a functional unit constituting a skeletal muscle, which can only contract and relax in response to changes in Ca+concentration. In order from the simple to the most complex, it builds structures corresponding to myofibrils, muscle fibers, muscle fiber bundles and the skeletal muscle. This distinctive hierarchical structure of skeletal muscles has been intensively studied in interdisciplinary research fields. In engineering, how the system efficiently controls a large number of sarcomeres to express continuous output force, is a point that has been focused. In this research, we propose a new decentralized control which is very simple but can manage many binary functional units by exploiting environmental noise. The validity of method is confirmed in both numerical simulation and a developed biologically inspired actuator.

This paper presents a direct teaching method for musculoskeletal robots driven by pneumatic artificial muscles (PAMs). In order to reproduce motions which are directly taught by a human, it is necessary to reproduce the lengths of PAMs because they geometrically determine the posture assumed by the robot. However, it is difficult to measure the lengths of PAMs because mounting length sensors is space-consuming. Additionally, estimating lengths is also difficult because it is required to know the intrinsic parameters of PAMs which are extremely difficult to measure for each muscle. In order to overcome the above problems, the proposed method calculates the desired internal pressures or the desired axial tensions of the PAMs under a specific constraint, which forces PAM's lengths in the reproducing phase to be similar to the lengths during the teaching phase. In this way, it is possible to reproduce the motion by controlling the internal pressures or the axial tensions instead of the lengths. The validity was confirmed through an experiment using a real musculoskeletal robot arm.

In recent years, musculoskeletal robots are being intensively studied to exploit the advantages of biological musculoskeletal systems for robot developments. In these robots, it is very important to assure engineering, biomechanical, and anatomical plausibility at the same time. However, these requirements are often in contradiction. Especially, in the human's shoulder complex, mimicking the glenohumeral joint and the scapulothoracic joint has been difficult because of the need to assure a wide range of movement and the joint's stability at the same time. In this paper, we propose mechanical structures to realize the functions of the glenohumeral joints and scapulothoracic joint. These structures were used to develop a musculoskeletal robot arm driven by pneumatic artificial muscles. In addition, in order to verify the feasibility of the robot arm, we present a dynamic motion in which the robot arm throws a ball by a simple control strategy.

Stochastic resonance (SR) is a phenomenon by which the immeasurable input signals of a non-linear system can be observed in the output signals by adding a non-zero level of noise. So far, this phenomenon has been intensively studied, but no methodology for its use in engineering applications has been established yet. To exploit SR in engineering, and, in particular, to determine the appropriate noise variance that optimizes the SR performance, remains an open problem. In this study, we propose a method, which exploits the non- linear correlation between outputs from a set of redundant sensors subject to different noise sources, to tune the noise variance. Because the proposed method does not require the input signal information, it allows the exploitation of SR in realistic engineering problems. The proposed method is validated by information theory and numerical simulations. Based on the results, we developed a tactile sensing system that utilizes the method. Experimental results demonstrate that the tactile sensing system can sense immeasurable signals which are smaller than the quantization error of the sensing system.

This paper describes a self-protective whole-body control method for humanoid robots. A set of postural reactions are used to create whole-body movements. A set of reactions is merged to cope with a general falling down direction, while allowing the upper limbs to contact safely with obstacles. The collision detection is achieved by force sensing. We verified that our method generates the self-protective motion in real time, and reduced the impact energy in multiple situations by simulator. We also verified that our systems works adequately in real-robot.

In order for humans and robots to engage in direct physical interaction several requirements have to be met. Among others, robots need to be able to adapt their behavior in order to facilitate the interaction with a human partner. This can be achieved using machine learning techniques. However, most machine learning scenarios to-date do not address the question of how learning can be achieved for tightly coupled, physical touch interactions between the learning agent and a human partner. This paper presents an example for such human in-the-loop learning scenarios and proposes a computationally cheap learning algorithm for this purpose. The efficiency of this method is evaluated in an experiment, where human care givers help an android robot to stand up.