Data-Driven Autoregressive Power Prediction for GTernal Robots in the Robotarium

Yassin Abdelmeguid, Ammar Hasan · 2026

Read Data-Driven Autoregressive Power Prediction for GTernal Robots in the Robotarium by Yassin Abdelmeguid, Ammar Hasan (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-13908.txt.

Introduction

Abdelmeguid and Hasan build a small neural predictor for power use in GTernal robots at the Georgia Tech Robotarium. Past power readings predict future use far better than motion data alone, with a lag-one correlation of 0.95. Their 7,041-parameter network reaches an R2 of 0.90 on held-out motion patterns. On seven real robots in a collision avoidance task, it scores 0.87 with no retraining. Each prediction takes 224 microseconds, fast enough for real-time control.

Catalog record

CREATORYassin Abdelmeguid, Ammar Hasan
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE18201
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.13908

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Data-Driven Autoregressive Power Prediction for GTernal Robots in the Robotarium Yassin Abdelmeguid, Ammar Hasan Abstract Energy-aware algorithms for multi-robot systems require accurate power consumption models, yet existing approaches rely on kinematic approximations that fail to capture the complex dynamics of real hardware. We present a lightweight autoregressive predictor for the GTernal mobile robot platform deployed in the Georgia Tech Robotarium. Through analysis of 48,000 samples collected across six motion trials, we discover that power consumption exhibits strong temporal autocorrelation ( ρ 1 = 0.95 \rho_{1}=0.95 ) that dominates kinematic effects. A 7,041-parameter multi-layer perceptron (MLP) achieves R 2 = 0.90 R^{2}=0.90 on held-out motion patterns by conditioning on recent power history, reaching the theoretical prediction ceiling imposed by measurement noise. Physical validation across seven robots in a collision avoidance scenario yields mean R 2 = 0.87 R^{2}=0.87 , demonstrating zero-shot transfer to unseen robots and behaviors. The predictor runs in 224 μ \mu s per inference, enabling real-time deployment at 150 × \times the platform’s 30 Hz control rate. We release the trained model and dataset to support energy-aware multi-robot algorithm development. Keywords mobile robotics \cdot energy prediction \cdot multi-robot systems \cdot power consumption \cdot Robotarium 1 Introduction Multi-robot coordination algorithms increasingly incorporate energy awareness as a first-class objective, whether for coverage control with battery constraints, task allocation under power budgets, or reinforcement learning with energy-shaped rewards. These algorithms require power consumption models that can predict energy expenditure from commanded actions, enabling planners and policies to reason about battery depletion before committing to trajectories. Simulators for multi-robot systems similarly need accurate energy models to train policies that transfer to physical hardware without experiencing unexpected battery failures. The dominant approach to energy modeling assumes power consumption follows directly from robot kinematics, typically through polynomial functions of velocity magnitude or motor commands ( Hou et al. 2019 ) . While intuitive, this formulation ignores the complex thermal dynamics of motors, transient behaviors during acceleration, wireless communication overhead, and other hardware effects that contribute significantly to instantaneous power draw ( Góra et al. 2024 ) . These unmodeled dynamics manifest as large prediction errors when kinematic models encounter real telemetry data. We present a data-driven energy predictor for the GTernal mobile robot platform ( Kim et al. 2024 ) deployed in the Georgia Tech Robotarium ( Wilson et al. 2020 ) . The Robotarium provides remote access to a swarm of differential-drive robots equipped with INA260 power monitors capable of measuring voltage, current, and power at 10 mW precision. This infrastructure enables systematic data collection across diverse motion patterns while capturing ground-truth power consumption at the 30 Hz command rate. Our key insight emerges from statistical analysis of the collected data. Power consumption exhibits remarkably strong temporal autocorrelation, with lag-1 correlation \rho_{1}=0.95 across all motion trials. This autocorrelation structure implies that recent power history contains far more predictive information than current kinematic state, consistent with first-order autoregressive (AR(1)) process dynamics where future values depend linearly on immediate past observations ( Triebe et al. 2019 ) . A model that observes the previous power reading can explain 90% of variance in the next reading, while velocity features alone explain only 16%. The dominance of autoregressive structure over kinematic features suggests that power dynamics are governed by latent hardware states including motor thermal effects, WiFi transients, and sensor duty cycles that persist across timesteps but remain invisible to kinematic models. We contribute the following. First, we characterize power consumption dynamics on physical GTernal robots through systematic data collection across six motion protocols totaling 26.5 minutes of operation. Second, we design and validate an autoregressive Multi-layer Perceptron (MLP) predictor that achieves R^{2}=0.90 at the theoretical noise ceiling using only 7,041 parameters. Third, we demonstrate zero-shot transfer to seven unseen robots executing collision avoidance behaviors, achieving mean R^{2}=0.87 without retraining. Fourth, we release the trained model, dataset, and deployment code at [https://github.com/y4ssn/robotarium-energy-predictor]. 2 Related Work Energy modeling for mobile robots has traditionally relied on physics-based formulations that decompose power consumption into kinetic, frictional, and resistive components ( Hou et al. 2019 ) . However, comprehensive reviews of energy prediction techniques highlight that such analytical models struggle to account for computational overhead, sensor duty cycles, and thermal dynamics inherent to autonomous platforms ( Góra et al. 2024 ) . Recent work has shifted toward data-driven approaches, with neural architectures demonstrating superior accuracy by implicitly learning unmodeled dynamics from telemetry ( Visca et al. 2022 ) . Autoregressive neural networks have proven effective for energy prediction in related domains. Non-linear autoregressive models with exogenous inputs (NARX) successfully forecast building energy consumption by capturing thermal inertia alongside external driving variables ( Ruiz et al. 2016 ) . Similar architectures have been applied to unmanned aerial vehicle flight dynamics, where data-driven temporal models circumvent the complexity of analytical aerodynamics ( Dong et al. 2024 ) . These cross-domain successes motivate our application of autoregressive modeling to mobile robot power prediction. Recent advances in transfer learning for time-series prediction demonstrate that models trained on diverse temporal data can generalize zero-shot to unseen domains ( Kamalov et al. 2024 ) , suggesting that autoregressive predictors may transfer across robot instances without retraining. 3 Data Collection We collected power telemetry from a single GTernal robot executing scripted motion sequences on the Robotarium testbed. The GTernal platform ( Kim et al. 2024 ) is a differential-drive robot with an 11 cm \times 9.5 cm footprint and maximum linear speed of approximately 26 cm/s. The robot is equipped with an INA260 power monitor capable of measuring instantaneous power consumption with 10 mW precision. The onboard Raspberry Pi communicates velocity commands to a Teensy 4.0 microcontroller at 30 Hz, which executes a 100 Hz proportional-integral-derivative (PID) loop for motor control. The power monitor is polled at 10 Hz by the microcontroller, with the most recent reading returned alongside each velocity command, yielding synchronized telemetry at the 30 Hz command rate. Access to real-time power telemetry was enabled through API extensions implemented by the Robotarium team specifically to support this research. To ensure the predictor generalizes across motion patterns rather than memorizing specific trajectories, we designed six complementary trials covering the operational envelope of the platform. • Structured Motion. Eight phases cycling through idle, forward-backward translation, lateral movement, circular arcs, sinusoidal paths, velocity ramps, step changes, and figure-eight patterns. • Goal Navigation. Point-to-point navigation with proportional control to randomly sampled goal positions. • Acceleration Variation. Controlled acceleration and deceleration profiles at multiple rates. • Extremes. Maximum and minimum velocities including rapid direction reversals. • Random Walk. Uniformly sampled velocity commands with smooth interpolation. • Steady State. Extended constant-velocity segments to capture thermal equilibrium. Each trial collected 8,000 samples over approximately 4.4 minutes, yielding a combined dataset of 48,000 samples spanning 26.5 minutes of continuous operation. Power consumption ranged from 3,000 to 8,450 mW with mean 3,652 mW and standard deviation 375 mW. Linear velocity commands spanned -0.15 to 0.16 m/s while angular velocity reached \pm 3.14 rad/s. 4 Power Consumption Analysis Before designing a predictor architecture, we analyzed the statistical structure of power consumption to identify the most informative features. Figure 1 shows the autocorrelation function of power consumption across all six trials. The lag-1 autocorrelation \rho_{1} ranges

Read the complete text