Dynamic Task Space Control Enables Soft Manipulators to Perform Real‐World Tasks

Abstract

An emerging alternative to conventional rigid robotics is the field of soft robotics. [1][2][3][4] Soft robots adopt the principle of physical artificial intelligence to achieve an inherently compliant and embodied robotic behavior. [5,6] In soft robotics, robots are fabricated mainly from functional soft materials [7][8][9][10] to achieve inherently adaptive and intelligent characteristics. The inherent compliance of soft robots enables various behaviors that were previously difficult for conventional, rigid-linked robots to achieve, such as navigating through tight environments or gentle gripping. [11][12][13] Continuum robots [14][15][16] have shown improved adaptability and safety when moving from rigid to soft actuator materials and support structures. However, it is challenging to model continuum robots' behavior due to the infinite dimensionality of their state space.There have been several approaches to construct a dynamic model for soft continuum manipulators, akin to the manipulator equation. [17,18] Godage et al. [19] derived the dynamic model of a continuum manipulator through integral Lagrangian formulation with the assumption of continuous mass distribution along the arm. However, the closed-form expressions of dynamic terms become complicated as the number of segments increases, making the model unpractical for using it in realtime control. In order to simplify the dynamic model, the mass distribution of each segment was approximated into a single lumped mass. Under this assumption, Falkenhahn et al. derived the dynamic model for a continuously bending manipulator (Festo Bionic Handling Assistant) where the lumped masses are assumed to be concentrated in the tips of the soft continuum sections. [20] With the addition of a dynamic model, they showed acceleration-level control methods in joint space. In a later work, Falkenhahn et al. explicitly considered valve dynamics to achieve higher model accuracy. [21] Curvature space methods can also be combined with inverse kinematic approaches to control real-world coordinates. [22,23] This combination was shown by Gong et al., [24] who used an underwater continuum arm to grab samples. Alternatively, one can use machine learning methods to obtain a model of the dynamics. Thuruthel et al. use a recurrent neural network to approximate the dynamics, and learn a controller. [25] However, the blackbox nature of neural networks is undesirable for control.Alternatively, it is possible to compute the dynamic parameters of a continuously bending soft body by using an augmented rigid body model to approximate its kinematic and dynamic characteristics. [26,27] This model supplements the piecewise constant curvature (PCC) model by adding a rigid link model and mass

SofToss: Learning to Throw Objects With a Soft Robot

“…Soft robots may be better suited for this task, thanks to their characteristics [7]: they can store potential elastic energy within their bodies and then transfer it to the held object before releasing it to perform the throw. This characteristic has been shown in [8]. In that work, the authors show that a soft robotic arm is potentially able to perform several tasks, such as the throwing one, by releasing the held object during a linear trajectory.…”

Soft robotics towards sustainable development goals and climate actions

“…Embodied and physical intelligence plays a crucial role in developing autonomous soft robots that can effectively acquire, analyze, and utilize large amounts of data to tackle complex tasks such as navigating difficult trajectories, changing shape, and manipulating unknown and unstructured objects in various environments (Fischer et al, 2023). EI can simplify control parameters, resulting in more energy-efficient and cost-effective operations, but it requires a careful mechanical and material design to consider the interaction with the environment while performing a task (Laschi, 2022).…”

Materializing Autonomy in Soft Robots across Scales

“…[78,82,83]), thereby sacrificing task‐capability in favor of a well‐understood control strategy. [ 78,84 ]…”

“…[78,82,83]), thereby sacrificing task-capability in favor of a well-understood control strategy. [78,84] Even when the assumptions of PCC-based methods are compatible with a given task, control architectures designed around PCC are not necessarily optimal. [85,86] PCC-based control can be adapted to account for a soft robot's interactions with the environment (e.g., being loaded when picking up or interacting with an external object, or during walking as shown in Figure 3vii).…”