Configurable Battery Model Brings Real Cell Physics to Robot Simulators

An open-source Gazebo Sim plugin can reproduce measured lithium-ion battery voltage with root-mean-square errors of 33.4–34.3 millivolts when its parameters come from matched test sessions, according to the software’s validation study. The chemistry-configurable equivalent-circuit model is intended to address a conspicuous gap in robotic simulation, where several popular drone frameworks either omit a battery altogether or use a linear voltage model that cannot represent the polarization, diffusion, and charging–discharging hysteresis found in real cells. The project’s own analysis of Aerostack2, Flightmare, and Gazebo Sim found no equivalent-circuit battery model in those surveyed systems, although this is a source-code inspection rather than a controlled comparison of every available simulator plugin.

The new plugin, gz_ecm_battery_plugin version 1.3.0, implements the established Enhanced Self-Correcting formulation developed by Plett, so it does not claim a new battery theory. Its contribution is to deliver that physics as a maintained, simulator-independent C++ core with a Gazebo Sim interface, a permissive MIT license, and nearly all model inputs declared in Simulation Description Format files. The open-circuit-voltage relationship is supplied as an interpolated list of state-of-charge points rather than compiled into the software, allowing a user to characterize another cell by editing a world file. The model also accepts an optional second resistance-capacitance branch and state-of-charge-dependent resistance curves, while preserving the shared names, components, topics, and state-message interface of Gazebo Sim’s built-in LinearBatteryPlugin.

At each simulation step, the core calculates terminal voltage from open-circuit voltage, instantaneous and dynamic hysteresis terms, and one or two polarization branches, while tracking state of charge by coulomb counting with configurable discharge efficiency. The branch-current update uses an exponential zero-order-hold discretization rather than explicit Euler integration. That choice gives the update a decay factor between zero and one for every positive time step, making it unconditionally stable even if a simulator uses a large physics interval; an explicit method could diverge when the interval exceeds twice a branch’s resistance-capacitance time constant. The implementation also clamps state of charge to the interval from zero to one, prevents negative terminal voltage, holds a disabled branch at zero current, and reports at load time whether the declared voltage curve can physically reach a cell’s cutoff. Fourteen unit tests cover the numerical core and spline interpolation, while a separate end-to-end test checks the full path from SDF configuration through the plugin to published telemetry.

Validation used three publicly available measured datasets spanning NCA and NMC chemistries, cylindrical formats of 18,650 and 21,700 cells, and three laboratory protocols. On the Panasonic 18650PF, LG 18650HG2, and Molicel INR-21700-P42 A, the equivalent-circuit model produced full-trace voltage root-mean-square errors of 34.3, 34.0, and 33.4 millivolts, respectively. The corresponding optimistically fitted linear baselines produced errors of 36.0, 56.2, and 105.5 millivolts. When both models were scored on the second half of each trace after the linear baseline was fitted on the first half, the equivalent-circuit errors were 44.2, 38.0, and 39.8 millivolts, compared with 76.2, 118.6, and 134.5 millivolts for the baseline. These results concern measured terminal voltage at the cell level under isothermal laboratory conditions, not flight endurance or the performance of a complete drone battery pack.

The source of the open-circuit-voltage curve proved more consequential than simply increasing the number of model components. In a controlled experiment on the Panasonic cell, the model, test trace, and fitted resistance and polarization parameters remained unchanged while only the origin of the voltage curve changed. Using a separate slow C/20 discharge recorded two months earlier produced an 85.7-millivolt error, whereas deriving the curve from rest periods within the same HPPC session produced 34.3 millivolts, a factor-of-2.5 difference. The earlier and later sessions implied capacities differing by 6.6 percent, demonstrating that a plausible textbook method for measuring open-circuit voltage can introduce a substantial mismatch when combined with pulse data from another session. Parameters must therefore be identified under conditions that resemble those in which they will be used, not merely transferred because they describe the same nominal cell.

Temperature produced a similarly large penalty. When the model was parameterized and evaluated separately at −20, −10, 0, 10, and 25 degrees Celsius, voltage root-mean-square errors remained between 34 and 55 millivolts. The fitted series resistance increased from 25.4 milliohms at 25 degrees Celsius to 87.9 milliohms at −20 degrees Celsius, while usable capacity fell from 2.773 to 2.182 ampere-hours, a 21 percent reduction. Applying parameters fitted at 25 degrees Celsius to −20-degree data produced a 159.7-millivolt error, compared with 55.3 millivolts after reparameterization at the colder temperature. Version 1.3.0 carries no thermal state and does not automatically adapt its parameters, so each simulated temperature requires an appropriately characterized parameter set.

A separate transfer test used six automotive-style dynamic profiles, including US06, UDDS, LA92, and HWFET, to challenge a parameterization derived solely from a 25-degree-Celsius HPPC session. Across 19.8 hours of discharge, with currents reaching 8.1 times the cell’s rated capacity and including regeneration, the equivalent-circuit model gave whole-profile errors of 23.8–70.2 millivolts, while the linear baseline gave 27.8–72.6 millivolts. The aggregate advantage was small, and disappeared when the characterization session’s capacity was used instead of the capacity in each scored profile. After each model’s constant mean error was removed, however, the equivalent-circuit model tracked the most rapidly changing current samples better on all six profiles, with errors of 25.2–59.6 millivolts versus 31.2–78.4 millivolts for the baseline. The result supports the model’s state-dependent transient structure, but the experiment used bench-tested automotive duty cycles rather than the hover-and-manoeuvre current pattern of a multirotor.

Computational overhead is small enough for use inside a physics loop. On an AMD Ryzen 5 7535HS processor, a model step took 32 nanoseconds with one polarization branch and scalar resistance, 38 nanoseconds with two branches, and 51 nanoseconds with two branches plus state-of-charge-dependent resistance splines. Even the heaviest configuration represented about 0.005 percent of a one-millisecond physics step. These measurements covered the simulator-independent core; the surrounding Gazebo component writes and optional telemetry publication add costs belonging to the simulator. An integration demonstration attached the plugin to the Aerostack2 racing-quadrotor model without editing the upstream vehicle file, and its published state settled at 3.961 volts and 4.670 amperes under an 18.5-watt load, consistent with the expected power-to-voltage ratio. That check demonstrated correct system wiring rather than physical accuracy.

The study also identifies a consequential unresolved failure mode: the model reads voltage too high near zero state of charge. The largest residuals were 568, 285, and 349 millivolts for the three cells, even though their overall errors were about 34 millivolts. Making series or polarization resistance depend on state of charge yielded negligible net improvement after capacity correction, and adding a second polarization branch improved overall error by at most 4.5 percent on the Panasonic cell and less than 1 percent on the Molicel without specifically reducing the low-state-of-charge error. Separate hysteresis curves could not be identified reliably from the available rest points, and attempting to fit them degraded the result. The residual therefore remains an evidence-backed limitation rather than a defect that can be dismissed as insufficient model order. Its effect is operationally important because a controller relying on terminal voltage as a proxy for remaining charge could perceive a safety margin that the cell does not have. Version 1.3.0 should therefore not be treated as validated for the final few percent of discharge, pack imbalance, changing cell chemistry, or mission-level abort decisions without further testing.

Subject of Research: Validation of a chemistry-configurable equivalent-circuit battery model for robot simulation

Article Title: gz_ecm_battery_plugin: A chemistry-configurable equivalent-circuit battery model plugin for gazebo sim

Article References: Ponguillo-Intriago, R., Vanegas-Guillén, O., Yturralde-Villagómez, J. C., & Vasquez-Cevallos, L. (2026). gz_ecm_battery_plugin: A chemistry-configurable equivalent-circuit battery model plugin for gazebo sim. SoftwareX, 36, Article 103143. https://doi.org/10.1016/j.softx.2026.103143

Image Credits: AI Generated

DOI: 10.1016/j.softx.2026.103143

Keywords: battery modeling, Gazebo Sim, robot simulation, lithium-ion cells, equivalent-circuit model, drone simulation, energy-aware planning, software validation, battery, plugin, chemistry-configurable, equivalent-circuit