
Learning whole-body motion for a 19-degrees-of-freedom (DoFs) humanoid robot using deep reinforcement learning (DRL) is challenging due to high kinematic redundancy, large joint-level search spaces, and the need to maintain dynamic balance. This article proposes a structured DRL framework that integrates task-space via-point trajectory learning with a pseudo-inverse of Jacobian matrix for redundancy resolution (PJRR) method. Instead of learning full joint trajectories, a soft actor-critic (SAC) agent predicts a compact set of task-space via-points, which are converted into C-2-continuous motions using composite cubic splines. Kinematic redundancy is resolved analytically through the PJRR, which operates separately from the learned policy and enforces a three-level task hierarchy: end-effector trajectory subtask, hip trajectory subtask, and posture regulation subtask. The framework is validated on the KONDO KHR-3HV humanoid platform (1.3 kg, 19-DoFs) across four pick-and-place configurations, evaluated over N=24 independent trials per motion. Energy-aware trajectories are generated with no zero moment point (ZMP) violations in simulation. A time-incentive reward formulation characterizes the energy-time trade-off, and a time-conditioned policy enables speed generalization at inference without retraining. Preliminary transferability to a scaled 23-DoFs platform is demonstrated in simulation.
Compliant, powered exoskeletons have the potential to assist users in various circumstances including manual labor, rehabilitation, and elderly assistance. Parallel elastic actuators (PEAs) can potentially increase the battery life and/or assistance level of these devices, or allow the use of smaller motors while retaining comparable assistive performance. These benefits could be especially pronounced for quasi-direct drive (QDD) actuators, given the high currents required to deliver meaningful torques. Here, we design and validate a spring-cam-follower PEA system to augment our existing QDD knee exoskeleton (M-BLUE) based on a novel design framework with an associated open-access matlab application. The device employs a mixture of custom and commercial hardware to allow fast, low-cost implementation of a wide range of nonlinear parallel stiffness profiles. We use optimization to generate a stiffness profile to minimize Joule heating losses of the knee exoskeleton under a fractional assistance controller across four weighted tasks: squat lifting, level walking, and ramp ascent/descent. Benchtop dynamometer experiments characterize the empirical spring-cam stiffness curve, confirm the (IR)-R-2 energy savings achieved during target tasks (e.g., 40.3% reduction during squat with 25% assistance torque), and demonstrate the system's capacity for accurate open-loop torque control. An N=1 able-bodied human subject experiment serves as in situ validation of the PEA module during all four optimization tasks, with maximum Joule energy savings of 43.22 J seen during squat lifting/lowering-a 30% improvement compared to no PEA module.
Perching by grasping or landing on a branch or surface enables aerial vehicles to extend operation time. Existing perching devices either require an actuator or only support loads in one direction. We present the KGripper, a novel passive perching mechanism that is able to resist loads that both press into and pull away from the vehicle while perched. The design is based on the Kresling origami pattern, which exhibits coupled axial compression and twisting. We design a bistable Kresling dipole that drives a linkage to grasp a cylindrical object. Pushing the gripper into the object causes it to close, and twisting the gripper against the object causes it to open. The 122.5 g KGripper is able to support a vehicle over 6 kg, giving it a strength-to-weight ratio of at least 49. Compression tests indicate that a quadrotor using the gripper must be able to provide 11 N of thrust beyond hover and 0.3 Nm of yaw torque. We demonstrate successful perching on a quadrotor with a total system weight of 1.5 kg. Across both quadrotor and robot arm tests, successful grasping occurs in 96% of trials with objects of diameters up to 58 mm.
The design freedom of additive manufacturing has the potential to increase the complexity, functionality, and mechanical behavior of compliant mechanisms compared to conventional manufacturing methods. However, this same design freedom can result in "trial-and-error" design iteration when designing compliant mechanisms for target behaviors. Trial-and-error design iteration is undesirable and inefficient, particularly when designing compliant mechanisms with nonlinear-elastic materials and large deformations. To address this problem, we develop a synthesis method for the inverse design of compliant mechanisms that uses numerical optimization and a library of building blocks that incorporate nonlinear-elastic materials and large deformations as geometric nonlinearity. The synthesis method works by selecting the type and geometry of the building blocks that will deflect to a target shape under a given load, or achieve a target stiffness response. The method's ability to use different building block geometries and incorporate nonlinear behaviors, such as the superelasticity of shape memory alloys, is shown to enhance the design of compliant mechanisms to reach target shapes or stiffness responses.
This work presents a kinematic synthesis methodology that is developed to implement a simple linear stiffness element into an existing single-degree-of-freedom hand rehabilitation system without altering its mobility in order to render the system mechanically capable of replicating spasticity behavior during low-velocity rehabilitation treatments. Due to the velocity-dependent nature of spasticity, replicating its effects during these treatments conducted below certain velocity thresholds, where sudden increases in patient muscle tone are avoided, becomes a critical design constraint of the study. This allows utilization of a simple linear stiffness element on the system to replicate the increasing force disturbance observed during hand-opening motion in physical therapies. Throughout the study, spasticity behavior is replicated as a linearly increasing flexion-promoting force perpendicular to the fingers as the hand extension angle increases. In order to achieve this mechanically via a simple stiffness element, a new trajectory is created by considering desired force limits and the actual end-effector trajectory of the rehabilitation system. In light of this, a new mechanism that is actuated by the end effector of the system is synthesized to follow the constructed trajectory with minimal error. Having two distinct tip trajectories on the system with linearly increasing distances during hand extension, desired force disturbances as spasticity behavior were achieved via an attached spring with a suitable stiffness constant. Once the prototype of the system was constructed, results were verified by using motion capture cameras through the operation of the system.