🤖 AI

A Continual Learning Framework for Adaptive Control of Modular Soft Robots

This paper proposes a continual learning framework that enables modular soft robots to adaptively control varying morphologies and learn module-specific dynamics without forgetting prior knowledge, as validated through both simulation and real-world experiments.

Original authors: Nilay Kushawaha, Muhammad Sunny Nazeer, Baljinder Singh Bal, Cecilia Laschi, Egidio Falotico

Published 2026-07-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nilay Kushawaha, Muhammad Sunny Nazeer, Baljinder Singh Bal, Cecilia Laschi, Egidio Falotico

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a robot made of soft, squishy segments, like a chain of connected sausages or a caterpillar. This is a Modular Soft Robot (MSR). These robots are amazing because they can squeeze into tight spaces and move safely around people, but they are incredibly hard to control. They bend, twist, and stretch in unpredictable ways, making it difficult to write a computer program that tells them exactly how to move.

Usually, if you add a new segment to the robot or take one off, the old computer program stops working. You have to throw away the old "brain" and train a brand new one from scratch. This is slow and wasteful.

This paper introduces a new "brain" called SMPL (Soft Modular Progressive Learning). Here is how it works, using simple analogies:

1. The "Ever-Expanding Library" (Continual Learning)

Think of the robot's controller as a student learning to drive.

Old Way: If the student learns to drive a sedan, and then you swap the car for a truck, the student has to forget everything about the sedan and start over in a driving school for trucks.
The SMPL Way: This new framework is like a student who keeps their driving manual for the sedan on the shelf. When they get the truck, they don't throw the sedan manual away. Instead, they add a new chapter to their notebook specifically for trucks. They can still drive the sedan perfectly (because the old chapter is there), and now they can drive the truck too.

In technical terms, the paper calls this Continual Learning. The robot learns new shapes (by adding or removing modules) without "forgetting" how to control the old shapes. It builds a library of knowledge, adding new books as the robot changes, rather than rewriting the whole library every time.

2. The "Specialized Team" (Distributed Control)

Imagine a large orchestra.

Old Way: One single conductor tries to tell every single violinist, drummer, and flutist exactly what to do at the same time. If the conductor gets tired or makes a mistake, the whole orchestra suffers.
The SMPL Way: This framework acts like a conductor who hires a specialized coach for each section. One coach teaches the violins, another teaches the drums, and so on.
- Each "coach" (sub-network) only focuses on their specific instrument (module).
- They talk to each other (lateral connections) to share tips, but they don't rely on one person to control everything.
- The Benefit: If a drumstick breaks (a module fails), you only need to retrain the drum coach. The violin coach keeps playing perfectly. This makes the system more robust and precise.

3. The "Safety Net" (Closed-Loop Control)

The paper also describes a safety mechanism. Imagine the robot is trying to draw a circle.

Open-Loop (No Safety Net): The robot guesses the commands to draw the circle. If it starts to drift off, it keeps drawing a crooked line because it doesn't know it's wrong.
Closed-Loop (SMPL's Approach): The robot has a "mirror" (a forward model) that predicts where it will be. It compares that prediction to where it should be. If there is a difference (an error), it immediately feeds that error back into its brain to correct the next move. It's like a person drawing a circle while constantly looking at the paper and adjusting their hand in real-time.

What Did They Prove?

The researchers tested this "brain" in two ways:

In a Computer Simulation: They built a virtual soft robot and kept adding segments to it (from 1 segment up to 5). The SMPL framework learned to control the 1-segment robot, then the 2-segment, and so on, without losing the ability to control the 1-segment version. It worked better than other standard AI methods.
On a Real Robot: They built a real, physical robot with three soft, air-filled segments. They tested it on a real table. The robot successfully followed complex paths (like circles, rectangles, and spirals) with high precision. The "team of coaches" approach (distributed control) was more accurate than the "single conductor" approach.

The "Smart Energy Saver" Experiment

Finally, they showed the robot could be lazy in a smart way.

The Test: They asked the robot to reach a specific point in the air.
The Trick: Instead of using all three segments to reach the point, the robot figured out it only needed the middle segment. It turned off the other two to save energy and computing power.
The Result: If the middle segment wasn't enough, it would "wake up" the first segment to help. It only used the minimum number of parts necessary to get the job done.

Summary

This paper presents a new way to teach soft robots how to move. Instead of forcing the robot to relearn everything whenever it changes shape, this system adds new skills without forgetting old ones. It also breaks the control job into smaller, specialized tasks, making the robot more accurate and resilient. It's like giving a soft robot a brain that grows with it, rather than one that has to be replaced every time the robot changes its outfit.

Technical Summary: A Continual Learning Framework for Adaptive Control of Modular Soft Robots

Problem Statement
Modular Soft Robots (MSRs) offer significant advantages in unstructured environments due to their compliance and reconfigurability. However, controlling MSRs presents unique challenges stemming from their nonlinear dynamics, hyper-redundancy, and complex material properties. Existing control strategies face two primary limitations:

Lack of Adaptability to Morphological Changes: Most controllers assume a fixed robot morphology. When modules are attached or detached, the controller typically requires retraining from scratch, failing to leverage previously acquired knowledge.
Scalability and Robustness in Fixed Configurations: For MSRs with fixed morphologies, conventional approaches often rely on a single centralized controller. This limits modularity, reduces scalability for robots with many modules, and creates a system sensitive to error propagation. Furthermore, if a specific module fails (e.g., due to a puncture), the entire centralized model may require retraining.

Methodology: Soft Modular Progressive Learning (SMPL)
The authors propose SMPL, a control framework based on Continual Learning (CL) and Progressive Neural Networks (PNN). The framework addresses the problem of catastrophic forgetting while enabling incremental learning and distributed control.

Architecture: The SMPL framework utilizes a progressive architecture where each new MSR configuration (or each module in a fixed configuration) is assigned a dedicated sub-network.
- Lateral Connections: To facilitate knowledge transfer, lateral connections are established between the current sub-network and all previously learned sub-networks. The parameters of existing sub-networks are frozen to preserve prior knowledge, while new sub-networks are trained with random initialization.
- Mathematical Formulation: The hidden representation of the $i$ -th layer in the $k$ -th sub-network combines a self-contribution from the current sub-network and a lateral contribution from previous sub-networks (Eq. 1).
Closed-Loop Training: The framework operates in a closed-loop configuration. An inverse model (the controller) predicts actuation commands based on the target pose and current task-space error. These commands are passed through a pre-trained, fixed Forward Dynamics Model (based on LSTM) to predict the resulting pose. The error between the predicted and target pose is fed back as an input to the inverse model to stabilize training and reduce task-space error.
Training Data: The system is trained using "motor babbling" data, where random actuation signals are applied to explore the workspace.
Experimental Scenarios:
- Exp1S (Incremental Morphology): The framework learns sequentially as modules are added (from 1 to 5 modules). It must predict actuation for the current configuration while retaining the ability to control previous configurations without retraining.
- Exp2S & Exp2R (Distributed Control): For a fixed three-module MSR, the framework employs a distributed approach where each sub-network learns the specific dynamics of one module. This allows for localized control and potential retraining of only the affected module if degradation occurs.

Key Contributions

Incremental Adaptation: The proposal of a CL-based framework that allows MSRs to adapt to new morphological configurations (module attachment/detachment) while preserving knowledge of previously learned structures.
Distributed Control for Fixed Morphologies: Demonstration that the same framework can learn module-specific dynamics for fixed-size MSRs, offering improved precision and scalability compared to centralized controllers.
Comprehensive Validation: Extensive evaluation through simulation (tendon-driven robot) and real-world experiments (three-module pneumatic arm) across multiple trajectory tracking tasks.
Adaptive Module Activation: A demonstration of the framework's ability to selectively activate only the necessary modules to reach a target, reducing computational overhead and energy consumption.

Results

Trajectory Tracking (Exp1S): The SMPL framework successfully tracked trajectories across five incremental configurations (1 to 5 modules). It demonstrated the ability to retain performance on previously learned configurations after training on new ones, effectively mitigating catastrophic forgetting. Closed-loop training consistently yielded lower position and orientation errors compared to open-loop settings.
Benchmarking (Exp2S & Exp2R): In both simulation and real-world experiments, SMPL outperformed benchmark models, including Multi-Layer Perceptrons (MLP), standard LSTM, VAE-LSTM, and Bi-LSTM.
- Simulation: SMPL achieved the lowest position error (3.52 mm) and orientation error (4.69°) on test babbling data.
- Real-World: On the physical robot, SMPL achieved a position error of 4.97 mm and orientation error of 1.75°, significantly outperforming the Bi-LSTM baseline (7.49 mm and 2.47°).
Adaptive Activation: In the reaching experiment, the framework dynamically activated modules (starting with the target module, then preceding, then subsequent) to minimize error. This selective activation successfully reached the target with reduced actuation compared to activating all modules simultaneously.

Significance and Claims
The paper claims that the SMPL framework provides a robust solution for the control of modular soft robots by addressing the dual challenges of morphological reconfigurability and scalability. By leveraging continual learning, the system avoids the computational cost of retraining from scratch when the robot's structure changes. Furthermore, the distributed control architecture enhances system robustness; if a specific module degrades or fails, only the corresponding sub-network requires retraining rather than the entire system. The authors position this work as a step toward more efficient, adaptive, and scalable soft robotic systems capable of operating in dynamic environments.

Limitations and Future Work
The authors acknowledge that the sequential learning process results in increased training time. Future work aims to integrate meta-learning principles to accelerate adaptation to new modules. Additionally, the authors plan to explore the framework's robustness against real-world actuator degradation (e.g., leakage) and extend the architecture to actuator-level modular control for finer precision and energy efficiency.