Toward generic control for soft robotic systems

TL;DR

This paper tackles the fragmentation of control methods in soft robotics by proposing a generic, safety-aware framework that embraces control compliance. It couples a learning-based dynamics model (Neural ODE), a sampling-based planner (SBMPC), and an adaptive safety filter (CBF) that coordinate via a surrogate-action representation, inspired by human motor control. The approach is validated across three distinct soft platforms—a tendon-driven arm, a soft fish, and a cyborg cockroach—demonstrating stable, safe, and cross-platform transferable behavior. The results suggest that exploiting compliance and high-level tendencies, rather than enforcing precise low-level commands, yields robust, generalizable soft-robot control with real-time capabilities and resilience to model degradation. This framework has potential to unify soft-robot control and accelerate deployment across diverse morphologies and environments.

Abstract

Soft robotics has advanced rapidly, yet its control methods remain fragmented: different morphologies and actuation schemes still require task-specific controllers, hindering theoretical integration and large-scale deployment. A generic control framework is therefore essential, and a key obstacle lies in the persistent use of rigid-body control logic, which relies on precise models and strict low-level execution. Such a paradigm is effective for rigid robots but fails for soft robots, where the ability to tolerate and exploit approximate action representations, i.e., control compliance, is the basis of robustness and adaptability rather than a disturbance to be eliminated. Control should thus shift from suppressing compliance to explicitly exploiting it. Human motor control exemplifies this principle: instead of computing exact dynamics or issuing detailed muscle-level commands, it expresses intention through high-level movement tendencies, while reflexes and biomechanical mechanisms autonomously resolve local details. This architecture enables robustness, flexibility, and cross-task generalization. Motivated by this insight, we propose a generic soft-robot control framework grounded in control compliance and validate it across robots with diverse morphologies and actuation mechanisms. The results demonstrate stable, safe, and cross-platform transferable behavior, indicating that embracing control compliance, rather than resisting it, may provide a widely applicable foundation for unified soft-robot control.

Figures (6)

• Figure 1: Generic control framework for soft robots. This study presents a generic and safety-guaranteed control framework that enables unified control of diverse soft robotic systems. The unified framework is applicable to soft robotic arms, bioinspired fish robots, cyborg insects, as well as other soft robots operating across various environments, demonstrating morphology-independent control.
• Figure 2: Overview of the proposed control framework inspired by human motor intelligence. a Overall architecture. The control framework consists of three core modules: a learning-based model, a sampling-based planner, and an adaptive safety filter. The learning-based model employs neural ordinary differential equation (Neural ODE) to approximate the intrinsic robot–environment dynamics; the sampling-based planner generates candidate control sequences via sampling-based model predictive control (SBMPC) formulation and predicts corresponding outputs; and the adaptive safety filter, grounded on control barrier function (CBF), selects the optimal predicted motion within the provably safe region and maps it back to real-time control signals through a reciprocal mapping. This structure mirrors the human motor control process (coarse modeling, rapid planning, and reflex-like safety regulation) to maintain both flexibility and stability under uncertainty. b Learning-based model. The Neural ODE captures dynamics of the soft robot and provides bounded prediction errors for motion forecasting, similar to how the human brain forms and refines internal representations of body–environment interactions through perception and experience. c Sampling-based planner. The SBMPC-based formulation explores and optimizes candidate motion sequences using importance sampling, achieving efficient planning of soft robotic motion, much as the human nervous system performs rapid sample-based motion planning in complicated environments. d Adaptive safety filter. The adaptive CBF defines a provably safe region for the soft robot and converts the selected optimal prediction into stable, executable control inputs, ensuring robust and safe behavior under external perturbations. This mechanism resembles how the human nervous system triggers reflexive corrections when motion approaches safety limits, much like a parent catching a falling child.
• Figure 3: Experimental validation on the soft robotic arm. The experiments include trajectory tracking and obstacle avoidance of the arm's end-effector.a Structure and actuation of the tendon-driven soft arm. b Square trajectory tracking with safety constraints. Top and side views are shown in (i) and (ii). The end-effector follows a square path within the defined safety margin (blue band) (iii). (iv) shows the temporal evolution of the tracking safety function (TSF), defined as the minimum distance to the boundary (see (v)). It indicates safe operation when within 0-10. As shown in (iv), TSF remains stable over time, with slight dips at corners but no violations. Boxplots in (vi) show consistent TSF distributions across ten trials, with all trajectories remaining within the safe region. c Obstacle-avoidance performance. Top and side views in (i) and (ii) show the arm adapting its motion to avoid a green obstacle, with corresponding trajectories in (iii). The temporal evolution of TSF and the avoidance safety function (ASF) is shown in (iv), where ASF is defined as the minimum distance to the obstacle (see (v)). It remains positive throughout, ensuring safe obstacle avoidance, when TSF is relaxed to satisfy ASF constraints. Boxplots in (vi) show stable TSF and ASF distributions, confirming the controller’s consistency and reliability.
• Figure 4: Experimental validation on a soft-bodied robotic fish navigating a confined aquatic environment.a Structure of the soft-bodied robotic fish propelled by a multi-segment flexible tail. b Navigation in an obstacle-filled environment. The setup with two obstacles is shown in (i). The robot followed a figure-eight trajectory over ten trials, as seen in (ii), successfully avoiding collisions. Snapshots in (iii) illustrated smooth swimming of a representative experiment trial 1. c Turning control through input modulation. The control input is defined as the angle between the body and first tail segment (i). A sinusoidal signal governs tail motion, and turning is achieved by modulating its bias; the input trajectory for trial 1 is shown in (ii). Boxplots in (iii) confirm that the required tail oscillation amplitudes remain below the designed safety limit. d Evaluation of obstacle distances. Distances to obstacle 1 (blue) and 2 (red) are defined in (i). In trial 1 (ii), both remain above the safety threshold with minima of 0.33 m and 0.16 m. Boxplots in (iii) confirm safety over ten trials.
• Figure 5: Experimental validation on a cyborg cockroach performing straight-line locomotion. a Structure and task setup of the cyborg cockroach.The composition of the cyborg cockroach is shown in (i). The cockroach follows a straight-line path within the defined safety margin (blue band) (ii). b Straight-line locomotion with safety constraints. Top and rear views are shown in (i) and (ii), with stimulation events labeled for left turns (purple) and right turns (cyan). The cockroach remains within the safety region throughout. c Comparison with continuous stimulation strategy. (i) shows trajectories under both algorithms. With the proposed event-triggered strategy, most trajectories stay within the safety region (blue band). Under continuous stimulation, trajectories often exceed this region due to habituation. Boxplots in (ii) show the $|Y|$ distribution under both algorithms. The proposed method yields a stable distribution with lower variance and a smaller mean. As the number of stimulations increases, habituation emerges in both cases, leading to a decline in safety ratio, as shown in (iii).