Toward generic control for soft robotic systems

Technical Summary: Toward Generic Control for Soft Robotic Systems

Problem Statement
Conventional robotics equates motor intelligence with the generation and execution of precise actions derived from accurate system models. While successful for rigid robots with well-defined dynamics, this paradigm fails for soft robotic systems. Soft robots exhibit high-dimensional, time-varying, and strongly nonlinear dynamics due to their compliant structures and continuous physical interaction with the environment. Consequently, accurate modeling is often impractical, and model-based controllers become fragile, sensitive to errors, and unable to generalize across different morphologies or tasks. The authors argue that the fundamental issue is a mismatch in problem formulation: attempting to prescribe detailed low-level action commands in soft systems is neither realistic nor necessary. Instead, control should focus on regulating the boundaries of behavioral feasibility rather than generating specific trajectories.

Methodology
The paper proposes a generic control framework that shifts the control paradigm from "action generation" to "behavior regulation." Inspired by human motor control—where the nervous system produces coarse intentions and intervenes primarily to suppress unsafe behaviors—the framework operates without relying on system-specific models or detailed controller structures. It consists of three cooperating layers:

1. Learning-based Model (Neural ODE):

   • Instead of deriving analytical physics, the system uses a Neural Ordinary Differential Equation (Neural ODE) to learn a coarse, physically consistent representation of robot-environment interactions from data.
   • The model predicts potential outcomes of actions and explores dynamic regions beyond the training distribution, capturing underlying physical regularities rather than exact data samples.
   • It provides a "surrogate action" representing a high-level movement tendency (e.g., "extend arm forward") rather than a final control command.

2. Sampling-based Planner (SBMPC):

   • This module performs stepwise decision-making similar to human path planning. It generates a large set of candidate motion sequences via importance sampling.
   • These sequences are propagated through the Neural ODE to predict outcomes.
   • Trajectories are evaluated based on cost functions (tracking error and safety) and weighted exponentially. The optimal sequence is selected, but only the first action is executed before the cycle repeats, allowing for rapid adaptation to environmental changes.

3. Adaptive Safety Filter (CBF):

   • To address uncertainties in the coarse model and environmental variations, an adaptive Control Barrier Function (CBF) acts as a reflex-like safety filter.
   • It defines a provably safe region. If the planner's optimal prediction ($f_{optimal}$) falls within this region, it is passed through. If it approaches or crosses the safety boundary, the filter intervenes instantly.
   • The filter selects the closest safe prediction ($f_{safe}$) from the candidate set and ensures the system remains within the admissible set, effectively correcting unsafe tendencies without requiring precise dynamic models.

4. Reciprocal Mapping:

   • A critical architectural component that bridges the gap between the "surrogate action" level (where planning and safety occur) and the executable control input level.
   • Since the Neural ODE is a black box and traditional CBFs require explicit dynamics, this module establishes a bidirectional mapping between the sampled candidate control inputs ($u$) and their corresponding predicted motions ($f$).
   • Once a safe motion $f_{safe}$ is selected, the mapping retrieves the specific control input $u_{safe}$ that generated it, ensuring consistency between the learned dynamics and the executed command.

Key Contributions

• Paradigm Shift: The paper proposes a shift from precision-centric control (generating exact actions) to feasibility-centric control (regulating behavioral boundaries), arguing that intelligence in complex soft systems arises from constraining unsafe behaviors rather than prescribing details.
• Model-Agnostic Framework: The framework achieves generalization across diverse soft robotic platforms (tendon-driven arms, soft fish, cyborg insects) without requiring system-specific modeling or controller tuning.
• Integration of Learning and Safety: It successfully combines a data-driven, coarse-grained learning model with formal safety guarantees (via adaptive CBFs), overcoming the incompatibility between black-box models and traditional safety-critical control.
• Reciprocal Mapping Mechanism: The introduction of a reciprocal mapping block allows the coordination of learning, planning, and safety modules through a shared intermediate representation, enabling the translation of high-level tendencies into low-level commands.

Experimental Results
The framework was validated on three distinct platforms with varying morphologies, actuation methods, and environments:

1. Tendon-driven Soft Arm: Successfully tracked a square trajectory and avoided a fixed obstacle in a planar workspace. The system maintained safety constraints with a minimum Tracking Safety Function (TSF) value of 0.70 across 10 trials, demonstrating robustness in high-compliance systems.
2. Soft-bodied Robotic Fish: Performed autonomous obstacle avoidance in a confined aquatic environment, navigating around two spherical obstacles. The fish maintained safe distances of 0.33 m and 0.16 m to the respective obstacles, achieving stable directional control with minimal input modulation over 10 trials.
3. Cyborg Cockroach: Guided a live Madagascar hissing cockroach along a straight path with a narrow safety corridor ($\pm 0.05$ m). The controller triggered interventions only when necessary, maintaining high safety ratios with fewer stimulations compared to continuous stimulation strategies, thereby reducing habituation risks in the biological agent.

Robustness and Adaptability:

• Long-term Performance: In a 3000-cycle experiment, the framework compensated for model degradation (due to material fatigue) by using the adaptive CBF to update safety constraints in real-time, maintaining performance where non-adaptive controllers failed.
• Physical Constraints: The system successfully executed tight turns for the robotic fish under strict input limits, outperforming PID controllers that suffered from integral windup or saturation issues.
• Real-time Control: The sampling-based architecture enabled real-time tracking of moving targets (up to 0.32 m/s) by efficiently exploring the solution space through parallelizable trajectory sampling, avoiding the computational bottlenecks of traditional MPC.

Significance and Claims
The authors claim that this work addresses a fundamental gap in soft robotics: the lack of a unified control principle that enables generalization beyond morphology-specific designs. By decoupling control logic from specific actuation layouts and focusing on high-level behavioral tendencies, the framework provides a principled explanation for consistent performance across heterogeneous systems.

The paper posits that soft robots do not require highly customized control designs for each individual system. Instead, explicitly leveraging inherent compliance and regulating behavioral boundaries can yield control strategies with stronger generalization capabilities. This perspective aims to advance the field from system-specific engineering solutions toward a more unified and transferable soft-robotic science. The authors acknowledge that while they have validated the framework on major classes of soft robots, exhaustive evaluation of all possible morphologies remains a future direction.