Physics-informed digital twin and onboard control of a brainbot for intelligent active matter

arXiv:2511.00384 · cond-mat.soft · Submitted 2025-11-01 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "Physics-informed digital twin and onboard control of a brainbot for intelligent active matter".

Mira: Here's the gist of the paper. Establishing adaptive particles that sense their state, anticipate their evolution, and compute control inputs onboard has been a major challenge in non-equilibrium physics This work introduces a physics-based model that describes the various trajectories executed by a brainbot by decoupling the experimentally-observed translational, spinning, and orbital components of the motion The approach provides a scalable platform for machine-learning-enabled multi-agent studies and lays the groundwork for intelligent, adaptive active matter.

Kai: Kinematic Modeling of Brainbot Motion The authors construct a physics-informed digital twin based on a kinematic model that reproduces measured trajectory statistics and generates long, statistically faithful synthetic trajectories This model is founded on the kinematics of an elliptical shape described by its geometric center position vector r(t) and orientation angle φ(t).

Mira: The relationship between the position of an arbitrary point r' and the brainbot's frame is given by Eq.

Kai: (4), which relates them using scale factors ρ1 and ρ2.

Paper summary: Kai: Looking at the title, "Physics-informed digital twin and onboard control of a brainbot for intelligent active matter," it seems to be about moving beyond just mimicking movement to actually giving the particle control <ref:2511.00384#pg1>.

Mira: Right. The authors are addressing that major challenge in non equilibrium physics, which is establishing adaptive particles that sense their state, anticipate their evolution, and compute control inputs onboard <ref:2511.00384#pg1>.

Lev: From an engineering standpoint, what this paper really does is provide a way to implement predictive control on a real bristlebot using TinyMPC because it's computationally light for microprocessors <ref:2511.00384#pg2>.

Kai: They use the kinematic model to derive virtual control inputs, mapping them to desired absolute velocity and angular velocity using those kinematic relations <ref:2511.00384#pg3>.

Mira: This whole setup means they've demonstrated the first ever on board predictive control on a bristlebot, which is pretty significant work for active matter research <ref:2511.00384#pg5>.

Lev: The paper notes that because of this kinematic framework, it’s computationally light enough to be suitable for onboard applications on microprocessors with low compute capability <ref:2511.00384#pg6>.

Kai: So, the implication is that having a physics informed digital twin allows researchers to train machine learning models for complex desired trajectories by generating large ensembles of numerical trajectories <ref:2511.00384#pg6>.

Mira: It also suggests this kinematic framework can help find more efficient algorithms of motion that lead to diffusion on a given time scale <ref:2511.00384#pg6>.

Lev: The limitation they state is that relying on this kinematic framework enables the demonstration of on board predictive control, but it doesn't solve the fundamental problem of designing better physical motions themselves <ref:2511.00384#pg6>.

Kai: So, to wrap up, "Physics-informed digital twin and onboard control of a brainbot for intelligent active matter" provides a computationally light framework that decomposes complex motion into linear, spinning, and orbital modes while enabling the first on board predictive control demonstration <ref:2511.00384#pg6>.

Conclusion: Kai: So, we've been looking at this paper about building a physics informed digital twin and using onboard control for a brainbot <ref:2511.00384#pg6>.

Mira: It’s about taking this complex motion of the brainbot, breaking it down into separate parts like spinning and orbital motion, and then creating a model that mimics it very closely <ref:2511.00384#pg2>.

Lev: The authors used a kinematic model to reproduce measured trajectory statistics and they built an offline parametric identification model to fit the real experimental data <ref:2511.00384#pg3>.

Kai: They then use this twin for onboard predictive control using TinyMPC which is designed to run on resource constrained microprocessors <ref:2511.00384#pg4>.

Mira: The main point here is that they can generate synthetic trajectories that look really similar to the experimental ones, even when you push the system through different driving protocols <ref:2511.00384#pg5>.

Lev: It shows that this approach allows for a lot of testing because you're not limited by just what you can physically run in a lab setup <ref:2511.00384#pg6>.

Kai: The conclusion is that this framework decomposes the motion into linear, spinning, and orbital components which helps us understand how these active particles move <ref:2511.00384#pg6>.

Mira: It gives us data on an infinite domain and long time periods which really overcomes some of the limitations we hit when we only have experimental snapshots <ref:2511.00384#pg6>.

Lev: This capability to generate large ensembles of numerical trajectories is pretty useful for training machine learning models to achieve complex desired trajectories <ref:2511.00384#pg6>.

Kai: So, this work really sets the groundwork for intelligent active matter by showing we can have adaptive particles that sense their state and anticipate what they need to do <ref:2511.00384#pg1>.

Mira: It moves us toward a future where these systems aren't just reacting to things, but actively computing control inputs on the fly <ref:2511.00384#pg6>.

Lev: What this means for the world is that we can finally demonstrate on board predictive control on a bristlebot <ref:2511.00384#pg6>.

I. Mammadli, P. Shrestha, J. Pande, * F. Novkoski, † S. Mohapatra, M. Noirhomme, A.-S. Smith

Institute for Theoretical Physics at FAU Erlangen-Nürberg University · Pattern Recognition Lab at FAU Erlangen-Nürberg University · Department of Physical and Natural Sciences at FLAME University Pune India · GRASP Institute of Physics B5a at University of LiǤge Belgium · Group for Computational Life Sciences at Ruǭger Bośkoviž Institute Zagreb Croatia

cond-mat.soft

Submitted: 2025-11-01

Updated: 2026-11-13

Journal ref: Phys. Rev. Research 8, 033164 (2026)

DOI: 10.1103/rd1t-h2v6

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 78/100

The gist: Establishing adaptive particles that sense their state, anticipate their evolution, and compute control inputs onboard has been a major challenge in non-equilibrium physics This work introduces a…

Key concepts

Kinematic Modeling
This involves creating a mathematical model based on the geometry of an elliptical shape (position vector and orientation angle) to describe how the brainbot moves. It quantifies motion by defining a curvature parameter ($\eta$) that measures how much the actual trajectory resembles spinning motion.
Digital Twin Construction
A virtual replica of the physical brainbot is built using a forward kinematic framework integrated numerically. This model is refined by fitting it to experimental data using gradient descent to accurately capture the bot's specific movement characteristics.
Model Predictive Control (MPC)
This control method uses the digital twin model to predict future states over a set time horizon. It calculates optimal control inputs—like desired velocity and angular velocity—to minimize a cost function that balances tracking accuracy against control effort, ensuring efficient onboard decision-making.

Terminology

Summary

Establishing adaptive particles that sense their state, anticipate their evolution, and compute control inputs onboard has been a major challenge in non-equilibrium physics This work introduces a physics-based model that describes the various trajectories executed by a brainbot by decoupling the experimentally-observed translational, spinning, and orbital components of the motion The approach provides a scalable platform for machine-learning-enabled multi-agent studies and lays the groundwork for intelligent, adaptive active matter. p. 0

Where the paper says it

This provides a proof of principle for a brainbot that senses its state, predicts its evolution, and computes control inputs onboard, unlike conventional active particles with fixed motility, thereby transforming the brainbot into an agentic physical entity. By integrating physical modeling, data-driven parameter identification, and control into a unified framework, our approach provides a scalable platform for machine-learning-enabled multi-agent studies and lays the groundwork for intelligent, adaptive active matter.

Kinematic Modeling of Brainbot Motion

The authors construct a physics-informed digital twin based on a kinematic model that reproduces measured trajectory statistics and generates long, statistically faithful synthetic trajectories This model is founded on the kinematics of an elliptical shape described by its geometric center position vector r(t) and orientation angle φ(t). The relationship between the position of an arbitrary point r' and the brainbot's frame is given by Eq. (4), which relates them using scale factors ρ1 and ρ2. p. 1

Where the paper says it

Let A1 and A2 be the semi-major and semi-minor axis lengths of the brainbot ellipse, and let nˆ1 and nˆ2 be the unit vectors, in the laboratory frame, along the major and minor axes of the ellipse, given by nˆ1(φ) = (cos φ, sin φ), (1) nˆ2(φ) = (− sin φ, cos φ). (2) Let r ′ denote an arbitrary point in the plane (inside or outside the ellipse) in the laboratory frame, and let its coordinates be ρ1A1 and ρ2A2 in the reference frame of the brainbot, where ρ1 and ρ2 are scale factors.

The curvature parameter η is introduced to quantify trajectory similarity to spinning motion, defined as η = ω r − rc / v. p. 2
Where the paper says it

At a given moment, the center of rotation of the brainbot may lie inside or outside the brainbot. With these definitions, we introduce a curvature parameter η to quantify the similarity of an observed trajectory to spinning motion η = |ω| |r − rc| |v| , (7) As discussed in our previous work [16], this parameter is useful to characterize the different trajectories, with values of η close to 0 indicating linear motion and values close to 1 indicating spinning motion.

Digital Twin Construction and Validation

The digital brainbot is built by first defining a forward kinematic framework that numerically integrates dr/dt based on the motion equations. To accurately represent the brainbot, an offline parametric identification model uses a gradient descent method to fit the model to experimental data, minimizing the least-squares error across trajectories. p. 5

Where the paper says it

These parameters must be inferred from the experimental data. To this end, we rely on a gradient descent method to minimize the overall least-squares error between the experiments and the kinematic framework across the trajectories, for given geometric and driving conditions.

This process involves extracting ω(t) by numerically differentiating φ(t) and then fitting the resulting distribution of ω values to a skewnormal distribution. p. 5
Where the paper says it

From these processed trajectories, ω(t) can be obtained by numerically differentiating φ(t) with respect to time. This sequence is used to generate a histogram of ω values for each trajectory across the entire run of the experiment and then fit the resulting distribution of ω to a skewnormal distribution.

The fidelity of the reconstruction is evaluated by generating artificial trajectories whose η descriptors quantitatively match experimental templates on identical time scales. p. 1
Where the paper says it

We also define an optimization procedure that fits the model to the experiments to generate relevant parameter sets that can describe the complex experimental trajectories with high accuracy, effectively building a physicsinformed digital twin. We systematically evaluate the fidelity of the reconstruction as a function of the number of Fourier modes employed, enabling the generation of artificial trajectories whose η descriptors quantitatively match the experimental templates on identical time scales.

Onboard Control Implementation

The kinematic model is employed to implement onboard predictive control using Model Predictive Control (MPC) [26]. TinyMPC is chosen for its computational efficiency on resource-constrained microprocessors. p. 7

Where the paper says it

Considering the computational constraints of the brainbot, TinyMPC [22] has been chosen for solving the control problem due to its well-balanced computational speed and memory footprint, and adaptability to small embedded systems. Moreover, its operation on the same class of microprocessors (ARM® Cortex series) as used in the brainbot has been demonstrated a priori [22, 27].

The MPC solver predicts the evolution of the state over a horizon N to find optimal control inputs that minimize a cost function involving tracking cost, control effort, and terminal cost. p. 7
Where the paper says it

The solver predicts the evolution of the state of the bot over a horizon of predefined length N via the linear system model explained later, searching for the sequence of control inputs {u (j) | j = k, . k+N −1} (k denoting the current time step), that minimizes the cost function, while constrained to the imposed state and control input bounds.

The virtual control input derived from TinyMPC is mapped to desired absolute velocity vcommand and angular velocity ωcommand using kinematic relations. p. 8
Where the paper says it

The virtual control input [a (k) x , a (k) y ] ⊤ provided by TinyMPC after solving the minimization problem represents the desired accelerations of the geometric center in the laboratory frame and is used to obtain updated virtual velocities vx and vy. These virtual velocities, along with the orientation φ transmitted by the positioning setup, are then mapped to a desired absolute value of the velocity vcommand of the geometric center of the bot and a desired angular velocity ωcommand about the center of rotation of the bot, through Eqs.

Synthetic Trajectory Generation

The fully trained digital twin generates ensembles of synthetic trajectories by drawing angular velocities ω from the experimentally-obtained skew-normal distribution. This allows for the generation of trajectories exhibiting distributions of the curvature parameter η that are similar to those found from experiments. p. 9

Where the paper says it

For instance, we have demonstrated how we can (i) pick random values of the angular velocity ω from its experimentally-obtained distribution, or (ii) incorporate different levels of noise in ω by including an increasing number of modes from the Fourier decomposition of the experimental ω. This allows us to generate simulation trajectories that exhibit distributions of the curvature parameter η that are similar to those found from the experiments.

The resulting digital twin is demonstrated to be a faithful and informative extension of the experimental brainbot, capable of reproducing experimental behavior across the full range of driving protocols. p. 1
Where the paper says it

Because the digital brainbot reproduces behavior accurately across the full range of η, spanning diverse driving intensities and geometries, it provides a powerful tool for optimizing task-specific driving protocols. More broadly, it enables systematic exploration of complex physical scenarios through a synergistic interplay between experiment and in silico realizations, made possible by the stability and fidelity of the digital twin established here.

Experimental Validation and Performance

The performance is validated by comparing four experimental and synthetic trajectories, showing good agreement between the simulated curves and experimental data for various driving protocols. The results confirm that the bot can systematically reproduce a lemniscate trajectory with a small error, validating the ability of the brainbot to solve a control problem onboard. p. 8

Where the paper says it

The results show that the bot is able to systematically reproduce a lemniscate trajectory with a small error, consistent between different realizations of the trajectory and in agreement with previous implementations of TinyMPC [22]. Besides this generally good performance, an overall asymmetry of the trajectory is also observed, likely due to the inherent mechanical imperfections of the bot in turning left or right, as well as to the approximation introduced by the linearization of control.

The framework demonstrates that even significant extrapolation in time keeps the statistical properties of the digital representation robust. p. 5
Where the paper says it

4, respectively), while the distribution of ω is maintained by construction. Even a significant extrapolation in time keeps the statistical properties of the

Conclusion and Future Directions

The kinematic framework successfully decomposes complex trajectories into basic modes of motion, including linear, spinning, and orbital components. The digital twin provides data on an infinite domain and long time periods to overcome experimental limitations. p. 9

Where the paper says it

This process effectively yields a functional digital replica of the brainbot. These digital twins are important for providing data both on an infinite domain and for long time periods, overcoming experimental limitations.

This framework can be used to find more efficient algorithms of motion that lead to diffusion on a given time scale. p. 9
Where the paper says it

In previous work, we have shown that brainbots can be programmed to execute diffusive motion on long time scales, while on short time scales they exhibit ballistic motion [16, 17]. The framework constructed here can be used to find more efficient algorithms of motion that lead to diffusion on a given time scale, without the need for a large number of experiments.

The ability to generate large ensembles of numerical trajectories is useful in training machine learning models for achieving complex desired trajectories. p. 9
Where the paper says it

Additionally, access to a large ensemble of numerical trajectories can prove useful in training machine learning models. Such approaches can learn from the generally stochastic statistics of brainbot trajectories to train neural networks to achieve complex desired trajectories, while incorporating specific interactions between multiple bots, which may possibly reveal novel dynamical regimes.

Relying on a kinematic framework has enabled the demonstration of the first-ever on-board predictive control on a bristlebot. p. 9
Where the paper says it

We plan to implement such machine learning methods in future work. Relying on a kinematic framework has enabled us to demonstrate the first-ever on-board predictive control on a bristlebot to efficiently track a reference trajectory.

The framework is computationally light, making it suitable for onboard applications on microprocessors with low compute capability. p. 9
Where the paper says it

Under our framework, a model predictive control approach that runs directly on the brainbot [22] [30] has been leveraged against the kinematic model to execute the tracking of a desired trajectory. The framework is computationally light, making it suitable for onboard applications on microprocessors with low compute capability.

Supplemental Details

The supplemental material provides the equivalence of constant velocity in the laboratory frame and sinusoidal velocity in the body-fixed frame. It also details the numerical integration scheme used for simulations using a Crank-Nicolson scheme. p. 4

Where the paper says it

A pseudo-code for the digital brainbot is given in Algorithm 1. Details of the numerical integration scheme used, including the rotation matrix M(¯ω∆t) and the mean velocity v¯c, are provided in Sec.

Furthermore, it presents a comparison of experimental and simulated trajectories as a function of Fourier modes included in the reconstruction. p. 1
Where the paper says it

We systematically evaluate the fidelity of the reconstruction as a function of the number of Fourier modes employed, enabling the generation of artificial trajectories whose η descriptors quantitatively match the experimental templates on identical time scales. Furthermore, we use our digital brainbot to generate artificial trajectories on time scales far beyond experimentally-accessible ones.

The cost matrices used in TinyMPC are also provided for validation. p. 8
Where the paper says it

The cost function consists of three terms, from left to right: the tracking cost, the control effort, and the terminal cost, with Q, P, and R being positive semidefinite cost matrices. The matrices Q and R are diagonal matrices with the values representing the weights of the respective costs, while the matrix P is a symmetric matrix generated by TinyMPC as a solution to the Riccati equation [28].

References

[1] C. Bechinger, R. p. 9

Where the paper says it

4000103267. Bechinger, R.

Di Leonardo, H. p. 9
Where the paper says it

Bechinger, R. Di Leonardo, H.

Löwen, C. Reichhardt, G. p. 9
Where the paper says it

Di Leonardo, H. Reichhardt, G.

Volpe, and G. p. 9
Where the paper says it

Reichhardt, G. Volpe, and G.

Volpe, Active particles in complex and crowded environments

[2] Z. T. Liu, Y. Shi, Y. Zhao, H. Chaté, X. qing Shi, and T. p. 9

Where the paper says it

88, 045006 (2016). qing Shi, and T.

H. Zhang, Activity waves and freestanding vortices in populations of subcritical quincke rollers

[3] R. R. Garza, N. Kyriakopoulos, Z. p. 9

Where the paper says it

118, e2104724118 (2021). Kyriakopoulos, Z.

M. Cenev, C. Rigoni, and J. p. 9
Where the paper says it

Kyriakopoulos, Z. Rigoni, and J.

V. I. Timonen, Magnetic quincke rollers with tunable single-particle dynamics and collective states

[4] V. Narayan, S. p. 9

Where the paper says it

9, eadh2522 (2023). Narayan, S.

Ramaswamy, and N. p. 9
Where the paper says it

Narayan, S. Ramaswamy, and N.

Menon, Long-lived giant number fluctuations in a swarming granular nematic

[5] C. Scholz, M. Engel, and T. p. 9

Where the paper says it

Menon, Long-lived giant number fluctuations in a swarming granular nematic, Science 317, 105 (2007). Engel, and T.

Pöschel, Rotating robots move collectively and self-organize

[6] C. Scholz, S. Jahanshahi, A. p. 9

Where the paper says it

9, 931 (2018). Jahanshahi, A.

Ldov, and H. p. 9
Where the paper says it

Jahanshahi, A. Ldov, and H.

Löwen, Inertial delay of self-propelled particles

[7] L. Caprini, D. Breoni, A. Ldov, C. Scholz, and H. Löwen, Dynamical clustering and wetting phenomena in inertial active matter

[8] Y. Chen and J. Zhang, Anomalous flocking in nonpolar granular brownian vibrators

[9] L. Caprini, B. Liebchen, and H. Löwen, Self-reverting vortices in chiral active matter

[10] O. Dauchot and V. Démery, Dynamics of a self-propelled particle in a harmonic trap

[11] P. Baconnier, D. Shohat, C. H. López, C. Coulais, V. Démery, G. Düring, and O. Dauchot, Selective and collective actuation in active solids

[12] O. Chor, A. Sohachi, R. Goerlich, E. Rosen, S. Rahav, and Y. Roichman, Many-body szilárd engine with giant number fluctuations

[13] L. Giomi, N. Hawley-Weld, and L. Mahadevan, Swarming, swirling and stasis in sequestered bristle-bots

[14] E. Zheng, M. Brandenbourger, L. Robinet, P. Schall, E. Lerner, and C. Coulais, Self-oscillation and synchronization transitions in elastoactive structures

[15] A. Altshuler, O. L. Bonomo, N. Gorohovsky, S. Marchini, E. Rosen, O. Tal-Friedman, S. Reuveni, and Y. Roichman, Environmental memory facilitates search with home returns

[16] M. Noirhomme, I. p. 0

Where the paper says it

Mohapatra,1 M. Noirhomme,4 A.

Mammadli, N. p. 0
Where the paper says it

Physics-informed digital twin and onboard control of a brainbot for intelligent active matter I. Mammadli,1 P.

Vanesse, J. Pande, A.-S. Smith, and N. Vandewalle, Brainbots as smart autonomous active particles with programmable motion

[17] F. Novkoski, M. p. 0

Where the paper says it

Pande,3, ∗ F. Novkoski,1, † S.

Mélard, M. Delens, F. Wéry, M. Noirhomme, J. p. 0
Where the paper says it

Mohapatra,1 M. Noirhomme,4 A.

Pande, A Maier, A. p. 0
Where the paper says it

Shrestha,2 J. Pande,3, ∗ F.

-S. Smith, and N Vandewalle Graspion: An open-source programmable brainbot for active matter research

[18] F. Becker, S. Boerner, V. Lysenko, I. Zeidis, and K. Zimmermann, On the mechanics of bristle-bots - modeling, simulation and experiments in ISR/Robotik 2014; 41st International Symposium on Robotics (2014) pp. 1–6

[19] T. Majewski, D. Szwedowicz, and M. Majewski, Locomotion of a mini bristle bristle robot with inertial excitation J. p. 1

Where the paper says it

More broadly, it enables systematic exploration of complex physical scenarios through a synergistic interplay between experiment and in silico realizations, made possible by the stability and fidelity of the digital twin established here. Finally, our model of the brainbot, which relies on kinematics and doesn’t require a complete resolution of the physics of the friction-based bristle-induced motion, is then employed to implement onboard predictive control of a single brainbot following a predefined trajectory.

Mech. Robot 9, 061008

[20] H. Löwen and B. p. 9

Where the paper says it

Ldov, and H. Löwen, Inertial delay of self-propelled particles, Nat.

Liebchen, Towards intelligent active particles in Artificial Intelligence and Intelligent Matter: Nanoscience, Soft Matter, Philosophy (Springer, 2026) pp. p. 0
Where the paper says it

This provides a proof of principle for a brainbot that senses its state, predicts its evolution, and computes control inputs onboard, unlike conventional active particles with fixed motility, thereby transforming the brainbot into an agentic physical entity. By integrating physical modeling, data-driven parameter identification, and control into a unified framework, our approach provides a scalable platform for machine-learning-enabled multi-agent studies and lays the groundwork for intelligent, adaptive active matter.

257–271

[21] I. Mahajan and B. Plancher, Robust and efficient embedded convex optimization through first-order adaptive caching in 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2025) pp. p. 1

Where the paper says it

By relying on a lightweight solver designed for small-scale embedded systems, we enable a single brainbot to autonomously compute its control inputs to track a prescribed reference trajectory and make real-time adjustments in case of deviations. To the best of our knowledge, this is the first demonstration of on-board control for a bristlebot, underscoring the novelty of the approach and its relevance to the experimental realization of intelligent active matter.

2984–2990

[22] K. Nguyen, S. Schoedel, A. Alavilli, B. Plancher, and Z. Manchester, Tinympc: Model-predictive control on resource-constrained microcontrollers in 2024 IEEE International Conference on Robotics and Automation (ICRA) (IEEE, 2024) pp. p. 1

Where the paper says it

Recent advances in the control algorithms tailored to resource-constrained microprocessors of the kind used on these bots have aided the development of such systems, enabling advanced control in real time [21–23]. In this work, we introduce a physics-based model that describes the various trajectories executed by a brainbot by decoupling the experimentally-observed translational, spinning, and orbital components of the motion.

1–7

[23] J. Wikner, D. Arnström, and D. Axehill, On real-time feasibility of high-rate mpc using an active-set method on nano-quadcopters (2026) arXiv:2603.09342 [math.OC]

[24] See Supplemental Material at [URL will be inserted by publisher] for equivalence of constant velocity in laboratory frame and sinusoidal velocity in body-fixed frame and the numerical integration scheme used in the simulations

[25] A. Savitzky and M. p. 5

Where the paper says it

Each trajectory lasts approximately 20 seconds and contains about 800 data points. To suppress high-frequency noise, we apply a Savitzky–Golay filter to the trajectories [25], following the procedure described in Ref.

J. E. Golay, Smoothing and differentiation of data by simplified least squares procedures. p. 5
Where the paper says it

These parameters must be inferred from the experimental data. To this end, we rely on a gradient descent method to minimize the overall least-squares error between the experiments and the kinematic framework across the trajectories, for given geometric and driving conditions.

, Anal. Chem. 36, 1627 (1964) https://doi.org/10.1021/ac60214a047

[26] B. Kouvaritakis and M. Cannon, Model predictive control, Switzerland: Springer International Publishing 38, 7 (2016)

[27] E. Çintaş and B. Özyer, A robust fault-tolerant control algorithm for gps-denied mini quadrotors using pidtinympc and visual-inertial odometry, Control Engineering Practice 169, 106779 (2026)

[28] R. Tedrake, Underactuated Robotics (2023)

[30] I. Mahajan, K. Nguyen, S. Schoedel, E. Nedumaran, M. Mata, B. Plancher, and Z. Manchester Graspion: An open-source programmable brainbot for active matter research arXiv preprint arXiv:2403. p. 0

Where the paper says it

Physics-informed digital twin and onboard control of a brainbot for intelligent active matter I. Mammadli,1 P.

18149

[31] B. Stellato, G. Banjac, P. Goulart, A. Bemporad, and S. Boyd, OSQP: an operator splitting solver for quadratic programs in Math Program Comput 12, 637 (2020)

[32] R. Verschueren, G. Frison, D. Kouzoupis, J. Frey, N. v Duijkeren, A Zanelli, B Novoselnik, T Albin, R Quirynen and M Diehl acados—a modular open-source framework for fast embedded optimal control in Math Program Comput 14, 147 (2022)

[33] J. Crank and P. Nicolson, A practical method for numerical evaluation of solutions of partial differential equations of the heat-conduction type Math. Proc. Camb. Philos. Soc. 43, 50–67 (1947)

Supplemental Material

The supplemental material demonstrates the equivalence of constant velocity in laboratory frame and sinusoidal velocity in body-fixed frame. It also provides the numerical integration scheme used for simulations using a Crank-Nicolson scheme. p. 4

Where the paper says it

A pseudo-code for the digital brainbot is given in Algorithm 1. Details of the numerical integration scheme used, including the rotation matrix M(¯ω∆t) and the mean velocity v¯c, are provided in Sec.

The comparison of experimental trajectories to simulated ones as a function of the number of Fourier modes used is given in Sec. p. 1
Where the paper says it

We systematically evaluate the fidelity of the reconstruction as a function of the number of Fourier modes employed, enabling the generation of artificial trajectories whose η descriptors quantitatively match the experimental templates on identical time scales. Furthermore, we use our digital brainbot to generate artificial trajectories on time scales far beyond experimentally-accessible ones.

S3. Finally, numerical values of the cost matrices used in TinyMPC are given in Sec. p. 4
Where the paper says it

A pseudo-code for the digital brainbot is given in Algorithm 1. Details of the numerical integration scheme used, including the rotation matrix M(¯ω∆t) and the mean velocity v¯c, are provided in Sec.

S4.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Model predictive control implementation can be enhanced by replacing linear doubleintegrator dynamics in TinyMPC with more complex numerical integration schemes, such as Crank-Nicolson, to better capture non-linear dynamics during trajectory prediction. This would improve tracking accuracy for complex experimental trajectories beyond those captured by the current linearized model.

  2. The offline parametric identification model can be improved by using a more sophisticated noise suppression technique than the Savitzky–Golay filter, such as a tailored Fourier mode selection strategy, to optimize the trade-off between descriptive accuracy and computational overhead. This would allow for more accurate extraction of parameters u from noisy experimental data.

  3. The digital brainbot can be extended to incorporate machine learning methods trained on the stochastic statistics of brainbot trajectories to achieve complex desired trajectories, such as super-diffusive or sub-diffusive motion. This capability allows the system to learn from experimental noise and generalize its behavior beyond purely kinematic models.

  4. The onboard control system can be made more robust by incorporating advanced optimal control toolchains like Acados instead of relying solely on OSQP for solving the quadratic program in TinyMPC. This would potentially mitigate issues arising from linearization and asymmetry in the kinematic problem, leading to higher precision in trajectory tracking.

Abstract

Establishing adaptive particles that sense their state, anticipate their evolution, and compute control inputs onboard has been a major challenge in non-equilibrium physics. We address this challenge by realizing an autonomous brainbot, building on a recently developed programmable bristlebot. First, we construct a physics-informed digital twin of the device, based on a kinematic model that reproduces measured trajectory statistics and generates long, statistically faithful synthetic trajectories. The kinematics forms the foundation for implementing onboard model predictive control (MPC), enabling autonomous trajectory tracking, demonstrated by accurate execution of a non-trivial target path. This provides a proof of principle for a brainbot that senses its state, predicts its evolution, and computes control inputs onboard, unlike conventional active particles with fixed motility, thereby transforming the brainbot into an agentic physical entity. By integrating physical modeling, data-driven parameter identification, and control into a unified framework, our approach provides a scalable platform for machine-learning-enabled multi-agent studies and lays the groundwork for intelligent, adaptive active matter.

Sources

Related papers