LunaDrive: A Delay-Compensated High-Voltage GaN FET-Based Motor Driver for Dynamic Robots with Flat BLDC Motors

Sota Yuzaki; Temma Suzuki; Hiromi Tada; Masanori Konishi; Kento Kawaharazuka; Kei Okada

cs.ROSep 18, 2026

LunaDrive: A Delay-Compensated High-Voltage GaN FET-Based Motor Driver for Dynamic Robots with Flat BLDC Motors

Authors: Sota Yuzaki, Temma Suzuki, Hiromi Tada, Masanori Konishi, Kento Kawaharazuka, Kei Okada

Organizations: Department of Mechano-Informatics, Graduate School of Information Science and Technology, The University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo, 113-8656, Japan

Abstract

The performance improvement of high-power flat BLDC motors has accelerated the development of dynamic robots. However, many commercially available servo motors assume operating voltages of 48 V or lower, which limits the maximum rotational speed. Dynamic robots require rapid energy generation, so this voltage constraint restricts motion performance. Therefore, driving motors beyond the rated voltage is desirable to increase the instantaneous maximum speed. On the other hand, semiconductor devices used in motor drivers have a trade-off between voltage rating and current capacity. Conventional drivers using Si MOSFETs have difficulty achieving both high-voltage and high-current operation. Although GaN FETs are promising, compact drivers that can be mounted on the rear side of flat BLDC motors remain limited. In this study, a motor driver for high-power flat BLDC motors using GaN FETs is developed. The effect of delay compensation in the high-speed region beyond the rated operating range is also investigated. In the experiments, under 96 V operation, a continuous current of 30 A was achieved with a heat sink attached. A peak current of 80 A and a maximum electrical frequency of 3110 Hz were confirmed. A high-speed load lifting experiment driven by a LiPo battery 24S (100 V) was also conducted, demonstrating applicability to dynamic robot operation.

Figures & tables

Fig. 1: LunaDrive enabling high-voltage, high-speed operation for dynamic robots

Fig. 2: LunaDrive composed of two boards: a controller board and a driver board

Fig. 3: Electronic block diagram of LunaDrive

Fig. 4: Schematic of GaN FET-based half-bridge circuit

Device	Type	On-Resistance	Size
		[ $\mathrm{m\SIUnitSymbolOhm}$ ]	[ $\mathrm{m}\mathrm{m}^{2}$ ]
FET-1	GaN FET	3	15
FET-2	MOSFET	2.5	116
FET-3	MOSFET	5.6	30

TABLE I: Performance comparison of 150-V power FETs

Manufacturer	Part Number	Nominal Cap.	Retention (DC Bias)	Rated Volt.	Package Size	Height	Areal Cap.
		[ $\mathrm{\SIUnitSymbolMicro F}$ ]	[%]	[ $\mathrm{V}$ ]		[ $\mathrm{mm}$ ]	[ $\mathrm{\SIUnitSymbolMicro F}\text{\,}{\mathrm{mm}}^{-2}$ ]
Murata	GRM32EC72A106KE05	10	25.1	100	3225M	2.5	0.078
TDK	C5750X7S2A226M280KB	22	45.5	100	5750M	3.1	0.088
TDK	CAA572X7S2A336M640LJ	33	47.0	100	6050M	6.9	0.129
TDK	C3225X7T2E334K200AA	0.33	69.3	250	3225M	2.0	0.029
TDK	C5750X7T2E225K250KA	2.2	58.5	250	5750M	2.5	0.045

TABLE II: Selected high-areal-capacitance MLCCs across package form factors and voltage ratings evaluated under practical DC bias conditions

Manufacturer	Part Number	Nominal Cap.	Rated Volt.	Package Size	Height	Quantity
		[ $\mathrm{\SIUnitSymbolMicro F}$ ]	[ $\mathrm{V}$ ]		[ $\mathrm{mm}$ ]	[-]
Murata	GRM32EC72A106KE05	10	100	3225M	2.5	76
TDK	C2012X7T2E104K125AA	0.1	250	2012M	1.25	18
TDK	CGA4J3X7R2E223K125AE	0.022	250	2012M	1.25	12

TABLE III: Capacitors mounted on the power supply line of LunaDrive

Fig. 5: Control block diagram of LunaDrive with delay compensation

Fig. 6: Overall system configuration

Fig. 7: High-power wire module using LunaDrive

Name	Value
Rated Voltage	$48\text{\,}\mathrm{V}$
No Load Speed	$5040\text{\,}\mathrm{r}\mathrm{p}\mathrm{m}$
Rated Torque	$1.3\text{\,}\mathrm{N}\mathrm{m}$
Peak Torque	$4\text{\,}\mathrm{N}\mathrm{m}$
Pole Pairs	21

TABLE IV: RO80 Frameless BLDC Motor Specifications

Fig. 9: Switching waveform of drain-to-source voltage at 10 A motor current

Voltage	Delay Compensation	Speed	Electrical Frequency
[V]		[rpm]	[Hz]
48	Disabled	4340	1520
48	Enabled	4340	1520
96	Disabled	5760	2020
96	Enabled	8890	3110

TABLE V: Results of the High-Speed Operation Experiment

Fig. 10: Current and speed responses at 96 V with and without delay compensation

Fig. 13: Current and board temperature when $80\text{\,}\mathrm{A}$ is commanded

Fig. 14: Temperature rise over time at different continuous current levels

Fig. 15: Overview of the high-speed load lifting experiment. The sequential snapshots show the lifting motion at 100 V.

Fig. 16: Current and speed responses during high-speed load lifting at 50 V and 100 V

Residual Physics-Informed Neural Networks for High-Fidelity BLDC Motor Modeling
Jul 10, 2026Haitham El-HussienyNeural Surrogate Modeling Physics-Informed ML
Department of Mechatronics and Robotics Engineering, Egypt-Japan University of Science and Technology (E-JUST), New Burg El-Arab City, Alexandria, Egypt
Battery-Aware Reinforcement Learning for Aggressive Quadrotor Flight
Sep 29, 2026Alejandro Sanchez Roncero, Olov Andersson, Petter OgrenRL Control Autonomous Drone Racing
Department of Robotics, Perception and Learning, School of Electrical Engineering and Computer Science, Royal Institute of Technology (KTH), SE-100 44 Stockholm, Sweden
PALTO: Physics-Informed Active Learning for Tri-Gate FinFET Design Optimization for Vertical Power Delivery
May 31, 2026Ayoub Sadeghi, Leonid Popryho, Inna Partin-Vaisband
University of Illinois Chicago Chicago, USA

Jul 10, 2026J/K move · Enter open · S save

Residual Physics-Informed Neural Networks for High-Fidelity BLDC Motor Modeling

Haitham El-Hussieny

Department of Mechatronics and Robotics Engineering, Egypt-Japan University of Science and Technology (E-JUST), New Burg El-Arab City, Alexandria, Egypt

Neural Surrogate Modeling Physics-Informed ML

Open paper AskPDF arXiv

Accurate dynamics modeling of Brushless DC (BLDC) motors is fundamental to high-performance robotic joint control. This paper presents a Physics-Informed Neural Network (PINN) with a deep residual (ResNet) backbone that learns a continuous-time surrogate of the full six-state BLDC motor dynamics. Given simulation time, applied three-phase voltages, and excitation parameters as inputs, the network directly predicts all motor state variables -- rotor angle, angular velocity, three-phase currents, and winding temperature -- while simultaneously satisfying the governing electromechanical and thermal ODEs through a composite physics-data loss. A curriculum scheduling strategy gradually activates the physics penalty to prevent premature convergence. Training runs are completed in under two minutes on a standard CPU. Crucially, once trained, PINN inference achieves latencies of 0.1--22, mu s per query, up to 118x faster than conventional ODE solvers, making it suitable for real-time observer and control applications.

Abstract

Explore similar work

Residual Physics-Informed Neural Networks for High-Fidelity BLDC Motor Modeling

Battery-Aware Reinforcement Learning for Aggressive Quadrotor Flight

PALTO: Physics-Informed Active Learning for Tri-Gate FinFET Design Optimization for Vertical Power Delivery