💻 computer science

From Digital to Physical Reservoir Computing: Co-Optimizing Soft Robotic Reservoirs via Dynamics Matching

This paper proposes and validates a co-optimization framework that aligns the dynamics of simulated soft robotic reservoirs with high-performing digital references through joint parameter tuning and state mapping, resulting in a 33.7% average performance improvement across various classification and forecasting tasks.

Original authors: Nicola Visentin, Maximilian Stölzle, Mariano Ramírez Montero, Francesco Braghin, Daniela Rus, Cosimo Della Santina

Published 2026-08-04

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Nicola Visentin, Maximilian Stölzle, Mariano Ramírez Montero, Francesco Braghin, Daniela Rus, Cosimo Della Santina

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 super-smart digital brain that can learn to predict the future or recognize patterns, but it's heavy, slow, and eats up a lot of electricity. Now, imagine you could swap that digital brain for a squishy, wiggly physical object—like a soft robot arm or a jellyfish—that naturally moves and reacts in complex ways. This is the world of Physical Reservoir Computing. Instead of simulating a brain on a computer, scientists use the real-world physics of a soft robot to do the heavy lifting. The robot's natural wiggles and bends act as a "reservoir" of memory, turning simple inputs into rich, complex patterns that a simple computer can easily read. The big question is: Can we just grab any soft robot off the shelf and use it, or do we need to tune it first? Usually, people just use the robot "as-is," hoping its natural chaos is good enough. But this paper asks a bold question: What if we could "pre-train" a soft robot to act exactly like a super-smart digital brain before we even turn it on?

The authors of this paper, a team from Delft University of Technology, MIT, and Politecnico di Milano, say that just using a soft robot as-is isn't the best way to go. They found that by co-optimizing the robot's physical shape, its control system, and a special translation map, they could make a simulated soft robot mimic the behavior of a high-performing digital brain almost perfectly. Think of it like this: If a digital reservoir is a master pianist playing a complex song, a standard soft robot is like a toddler banging on the keys. The researchers didn't just tell the toddler to try harder; they actually reshaped the toddler's hands, gave them a new set of instructions, and built a translator so that the toddler's banging sounded exactly like the master pianist's song.

Here is how they did it and what they discovered.

The Problem: The "As-Is" Robot vs. The Digital Dream

In the world of Reservoir Computing, you have a fixed, non-linear system (the reservoir) that takes in data and expands it into a high-dimensional state. In a digital computer, you can design this system perfectly. In a physical robot, the "reservoir" is the robot's own body. The catch? Physical robots are often messy. Their natural movements might not be the best kind of messy for computing.

Usually, researchers take a soft robot, hook it up to a computer, and hope for the best. They treat the robot's body as a fixed, unchangeable object. The authors argue this is a missed opportunity. They wanted to know: Can we tune the robot's body and its controller to match a specific, high-performing digital brain?

The Solution: The "Digital Twin" Training

The team created a training method where a "digital reference" (a very good digital brain) acts as the teacher, and the "physical reservoir" (a simulated soft robot) acts as the student.

The Teacher: They used a digital model called a Random Oscillators Network (RON). Imagine a bunch of springs and weights connected together, vibrating in a complex, chaotic dance. This digital dance is excellent at remembering past inputs and predicting the future.
The Student: They used a simulated soft robot made of "Piecewise Constant Strain" segments—basically, a digital model of a flexible arm that bends and twists like a real one.
The Magic Link: They didn't just tell the robot to copy the teacher's answers. They told it to copy the teacher's movements at the level of acceleration. They built a special "translator" (a mathematical map) that converts the digital brain's coordinates into the robot's physical coordinates.
The Optimization: Using a powerful computer, they simultaneously tweaked three things:
- The robot's physical parameters (how stiff or heavy it is).
- The controller (how the robot moves in response to inputs).
- The translator map (how to speak between the digital and physical languages).

They did this by minimizing the "error" between the digital teacher's acceleration and the physical student's acceleration. Crucially, they did this without running the robot through every single task first. They just matched the physics of the movement.

The Results: A Squishy Robot That Thinks Like a Digital Brain

The team tested this in a computer simulation (not on real hardware yet) using four different tasks:

sMNIST: Recognizing handwritten digits in a sequence.
ADIAC: Classifying time-series data.
Mackey-Glass & Lorenz96: Predicting chaotic, weather-like time series.

They tested robots of different sizes (reservoir dimensions of 6, 9, 12, and 15).

The Big Win:
When they compared their "co-optimized" robot to a standard "as-is" robot, the optimized one was a huge winner.

Across all tasks and sizes, the optimized reservoirs showed a mean relative improvement of 33.7% compared to unoptimized robots.
In terms of matching the digital teacher's movements, optimizing the robot's physical shape (morphology) was key. When they kept the robot's shape fixed and only tweaked the controller, the error was much higher. The full optimization (changing the shape and the controller) reduced the error by massive margins: 90.2% for sMNIST, 61.6% for ADIAC, 77.1% for Mackey-Glass, and 86.4% for Lorenz96.
The optimized robots retained 90.9% of the performance of the original digital brain.

The Catch (and the Confidence):
It is important to note that these results are simulations. The "robots" existed only in code. The authors are very clear that this is a "proof of concept." They didn't build a physical robot and test it in the real world yet.

They used small reservoirs (dimensions 6 to 15) because simulating larger, more complex robots is computationally expensive.
They found that as the reservoir got bigger, it became harder to match the digital brain perfectly, but the optimized robots still outperformed the unoptimized ones significantly.
They explicitly ruled out the idea that you can just use a random soft robot and get great results; the "as-is" approach was consistently worse.

Why This Matters

This paper suggests that we don't have to accept the limitations of a physical robot's natural shape. Instead of just hoping a soft robot is "good enough," we can mathematically design it to be a perfect partner for a specific type of computing.

The authors conclude that while their method works beautifully in simulation, the next step is to see if it works on real, wobbly, noisy hardware. They also point out that for this to be truly useful, we need to figure out how to handle the fact that real robots aren't perfect copies of their digital models. But for now, they have shown that with the right "pre-training," a soft robot can be coaxed into thinking almost exactly like a super-smart digital brain, opening the door to faster, more efficient, and more embodied artificial intelligence.

Technical Summary: From Digital to Physical Reservoir Computing via Dynamics Matching

Problem Statement
Physical Reservoir Computing (PRC) leverages the compliant, nonlinear dynamics of physical substrates (such as soft robots) to process temporal data, offering potential benefits in energy efficiency and real-time embodied processing. However, current PRC implementations often suffer from performance gaps compared to digital reservoirs. This disparity arises because physical reservoirs are frequently deployed "as-is" without pretraining or co-optimization, whereas digital reservoirs are explicitly tuned for Reservoir Computing (RC) tasks. Existing methods for improving reservoirs are often architecture-specific, rely on hand-tuned heuristics, or require repeated downstream task evaluations, making them computationally expensive and indirect for physical systems. There is a lack of a tractable method to design physical reservoirs whose intrinsic dynamics are inherently useful for RC without optimizing through every specific downstream task.

Methodology
The authors propose a dynamics-level pretraining framework where a physical reservoir is co-optimized to approximate the vector field of a high-performing digital reference reservoir. The approach avoids task-specific optimization during the pretraining phase, instead focusing on matching the underlying dynamics.

Key components of the formulation include:

Digital Reference: A second-order digital reservoir (specifically a Random Oscillators Network, or RON) serves as the target dynamics, defined by $\ddot{y} = f_{ref}(y, \dot{y}, u)$ .
Differentiable Physical Model: The physical reservoir (simulated soft robots) is modeled as a differentiable second-order system: $\ddot{Q} = f_{phys}(Q, \dot{Q}, T; \theta_{phys})$ , where $Q$ represents generalized coordinates and $T$ is actuation.
Co-Optimization Variables: The method jointly optimizes three sets of parameters:
- Physical Parameters ( $\theta_{phys}$ ): Morphological properties of the soft robots (e.g., stiffness, damping).
- State Map ( $\theta_{map}$ ): A diffeomorphic map $\phi$ that transforms the digital reference state to the physical state ( $Q = \phi(y)$ ) and its inverse.
- Controller ( $\theta_{contr}$ ): Feedforward and feedback controllers that generate actuation $T$ based on the input and state.
Optimization Objective: The system minimizes an acceleration-level equation-error loss. Instead of rolling out trajectories over time (which requires temporal integration), the loss compares the target digital acceleration $\ddot{y}$ with the mapped physical acceleration $\hat{\ddot{y}}$ . This allows for gradient-based optimization without simulating the full time evolution of the system.
Implementation: The authors instantiate this using simulated planar continuum soft robots (modeled via Piecewise Constant Strain dynamics) and a Random Oscillators Network as the reference. Optimization is performed using parallel multi-start gradient descent (Adam).

Key Contributions

Dynamics-Level Pretraining Perspective: The paper introduces a paradigm where physical reservoirs are co-optimized against structurally compatible, high-performing digital reference dynamics, rather than being treated as fixed or selected arbitrarily.
Formulation for Second-Order Systems: A specific mathematical framework is provided for differentiable second-order systems, integrating a diffeomorphic state map, feedforward-feedback control, and an acceleration-level objective that avoids temporal integration.
Proof of Concept: The authors demonstrate the feasibility of this approach using simulated soft robotic reservoirs, showing that the method can effectively match reference dynamics and improve downstream task performance.

Results
The authors evaluated the optimized reservoirs on four tasks: sequential MNIST (sMNIST) and ADIAC (classification), and Mackey–Glass and Lorenz96 (forecasting), across four reservoir dimensions ( $n_y \in \{6, 9, 12, 15\}$ ).

Dynamics Matching: Compared to a baseline where only the map and controller were optimized (with fixed robot morphology), the "full pretraining" (co-optimizing morphology, map, and controller) reduced the test Root Mean Square Error (RMSE) between target and physical accelerations by an average of 68.8% across all tasks (ranging from 61.6% to 90.2% depending on the task).
Downstream Performance: When inserted into a standard PRC pipeline (where only the readout layer is trained), the fully pretrained physical reservoirs achieved a mean relative improvement of 33.7% over unoptimized "as-is" physical reservoirs.
Comparison to Digital Reference: The optimized physical reservoirs retained approximately 90.9% of the performance of the digital RON reference on average.
Stability: The full pretraining approach demonstrated lower sensitivity to initialization (smaller standard deviations) compared to baselines with fixed morphology, which occasionally resulted in diverging simulations.

Significance and Claims
The paper claims to demonstrate the feasibility of dynamics-level co-optimization for simulated soft robotic reservoirs. The authors argue that their approach bridges the gap between digital and physical RC by providing a tractable way to design physical substrates that are inherently "RC-ready" without needing to optimize for every specific downstream task.

The authors remain modest regarding the scope of their results:

The experiments are conducted in simulation using reduced-dimensional reservoirs ( $n_y \leq 15$ ) to ensure computational tractability, which limits the absolute task performance compared to state-of-the-art digital RC.
The method is presented as a proof of concept. The authors acknowledge that future work must address hardware validation, the incorporation of unmodeled dynamics and sensor noise, and the scaling of the method to higher-dimensional physical systems.
They note that while the physical reservoir reduces digital computation during inference, the pretraining phase still relies on digital computation for the controllers and mapping, though this cost scales linearly with reservoir dimension once the map is absorbed into the readout.

The Problem: The "As-Is" Robot vs. The Digital Dream

The Solution: The "Digital Twin" Training

The Results: A Squishy Robot That Thinks Like a Digital Brain

Why This Matters

Technical Summary: From Digital to Physical Reservoir Computing via Dynamics Matching

More like this