1. Spatial Constraint Problem in Multi-Axis Drive

Designers of humanoid robots and robot hands must fit motors, reducers, and drive electronics into a space no larger than a single finger joint. In the traditional approach of using independent drive boards for each axis, current sensors, control MCUs, gate drivers, and communication circuits are replicated for every axis. This results not only in increased board area and component costs, but also accumulated heat dissipation in the confined space, potentially leading to malfunction and shortened lifespan. Therefore, the essence of the problem is not merely miniaturization, but achieving miniaturization, heat management, and reliability simultaneously.

STMicroelectronics has expanded its low-voltage (LV) servo drive lineup from single-axis solutions toward dual-axis, high-power, and high-integration directions. From service robots, robotic lawnmowers, AGVs, and machine tool solutions, in 2025 a 2x100 W robot joint dual FOC solution based on STSPIN32G4 directly targeting robot joints has been added. The roadmap in Figure 1 illustrates this development. This white paper follows this latest solution, examining how the design choices address the spatial constraints described above.

2. Structure of 150 W Ultra-Compact Dual Servo Drive

The question from the previous section leads to how to drive two axes on a single board. The small solution for robot joints controls two motors with FOC at a 150 W rated power on a single board measuring 4 x 4.5 cm. At the center of control is the STSPIN32G4, which integrates controller and gate driver, while the second inverter is managed by the STDRIVE101 gate driver. The inverter comprises six STL50DN6F7 MOSFETs, with STL20N6F7 separately placed for braking. Current signals are conditioned by TSV771ILT operational amplifiers. According to the presentation materials, without the integration degree of STSPIN32G4, the control section PCB area would have doubled, and without the patented technologies explained later, the inverter section area would also have doubled. Figure 2 shows this configuration in block diagram form.

In terms of connectivity, CAN and MODBUS are supported, with encoder interfaces accommodating ABZ and Hall sensors, as well as SPI encoders. Under conditions of driving both motors with FOC, CPU load stands at approximately 60% (detailed measurement conditions in section 5). Driving two axes in complete vector control while maintaining approximately 40% computational headroom summarizes this board's design objective. This headroom becomes the foundation for additional functions such as communication processing and predictive diagnostics.

3. Shared Sensing and Thermal Distribution Layout: Two Patented Technologies

How then does the circuit for two axes fit into single-axis level area? The answer lies in two patented technologies. The first is the multi-motor shared current sensing network (US Patent US 11,757,345 B2). Conventionally, current sensors and associated circuitry that must be placed separately for each motor are designed to be shared among multiple motors. As a result, current sensors, sensing networks, PCB area, and MCU pin allocation are all reduced, making this the decisive factor enabling the 4 x 4.5 cm dual board. However, the shared structure is not without cost. According to technical Q&A, the limiting factor for current sensing is PWM duty cycle, which exists regardless of acceleration or deceleration. Therefore, motor control software (MCSDK) manages this constraint through functions that correct ADC sample timing and limit duty cycle accordingly.

The second is a patent on junction temperature measurement and PCB layout for dual motors (US 12,264,976 B2). When two motors operate together on a narrow board, heat dissipation can become uneven. This patent uses a cross-placement layout of switching devices in the two inverters to balance thermal distribution, and provides precise overheat protection through direct junction temperature measurement. As a result, system lifespan and reliability are improved. Figure 3 shows the principle of both technologies and thermal distribution differences compared to conventional layouts. In summary, the choice to share sensors, layout, and controller was the key to solving both area and heat dissipation constraints.

4. System Configuration and Reconfiguration Flexibility

Miniaturizing a single board loses value if the overall robot wiring remains complex. This solution assumes a distributed architecture where an upper motion control board issues commands to each joint's dual drive via CAN bus, with each drive managing vector control of two motors at the joint level. Since a single central controller manages multiple joints over a single bus, wiring is simplified and system expansion is straightforward when increasing joint count. According to technical Q&A, the CAN shown in the diagram is merely an example; the interface can be customized to other communication protocols such as RS-485.

Hardware reconfiguration is also supported. Depending on application requirements, if one motor requires greater output, inverters can be paralleled via jumper resistors R4, R5, R6 to convert to a high-output single motor drive. Selection between dual drive and high-output single drive is achieved through jumpers alone, without hardware redesign. However, in parallel configuration, the capability to drive two motors independently is sacrificed, so the trade-off between axis count and power per axis must be decided at the design stage. Figure 4 shows this parallel connection structure. On the communication side, the solution is compatible with MIT CAN protocol widely used in robot actuator applications and supports MIT motion control mode, enabling adoption into existing robot control environments using that protocol without additional development.

5. Performance Verification: MCU Load, Current Loop Bandwidth, Overload Tolerance

Does the shared structure actually handle control of two axes? The first verification is MCU load measurement. Measurement results for two control configurations on the small board for robot joints are shown in Table 1. Even when increasing the current loop from 25 kHz to 40 kHz to expand control bandwidth, maximum load during dual FOC operation remains at 60%. As a result, approximately 40% computational headroom remains, providing the basis for embedding communication processing and predictive diagnostics functions on the same MCU. According to technical Q&A, 40 kHz is not the upper limit of this configuration; MCSDK settings support up to 100 kHz, but practical use is difficult due to CPU load limitations.

The second verification is current loop bandwidth. Based on 16 kHz PWM, frequency sweep from 1 Hz to 4000 Hz showed the frequency at which phase lag reaches 90 degrees to be 2952 Hz. This demonstrates that fast current control required for applications like robot joints, which need rapid torque response, is achievable. The third is overload tolerance. Under single motor, rated current 3.2 A (RMS), and MOSFET maximum temperature limit of 100 °C conditions, the sustainable duration curve versus overcurrent ratio was measured. A typical thermal performance curve was confirmed where shorter duration is sustained at high overcurrents and longer duration near rated current, demonstrating the thermal distribution layout of section 3 contributes to this overload tolerance. Figure 5 shows these performance test results.

6. High-Power Expansion: 2x1 kW EtherCAT Dual Servo and 3 kW Reference Design

Is the same design philosophy maintained at higher power levels? The industrial dual servo drive supports 1 kW per axis, totaling 2 kW at 48 V rated, and drives two motors using the same single STSPIN32G4 and STDRIVE101 as the compact board. The rated current per axis reaches a maximum of 20 A (RMS), with the shared current sensing network patent applied identically. Both incremental and absolute encoders are supported, and communication can be selected between EtherCAT or CAN. The layout is optimized for reliability even under unbalanced load conditions where the two axes have different loads. On a 13 x 13 cm board, communication section, main control section, two inverter power sections, auxiliary power, and encoder interface are arranged by region, designed as SIL2 compliant including STO. The block diagram in Figure 6 shows this structure and daisy-chain expansion configuration. According to technical Q&A, the daisy-chain is EtherCAT-based with independent PHY at each segment, processes frames on-the-fly, has no endpoint concept, and as a result multiple joint drives can be connected in series over a single communication line, simplifying wiring.

6.1 Firmware Structure

The firmware follows a layered architecture where CiA402 drive profile and EtherCAT protocol stack are built on top of MCSDK. ST provides industrial fieldbus communication plug-ins to MCSDK to handle interworking with EtherCAT stack, and the CiA402 profile establishes itself as a unified motor drive interface. Therefore, even if the lower communication environment changes, upper application code can be maintained without modification, and developers can focus on motor control logic rather than lower-level communication integration. The plug-in includes an improved scheduler providing synchronized task timing, and encoder alignment via microshake method enables initial alignment without index signal or Hall sensor.

6.2 3 kW Single-Axis Reference Design

For cases where greater output is needed on a single axis, a reference design is also available. Supporting 3 kW with active cooling and 2 kW with passive cooling on a 8 x 5 cm board, optimized for 48 V applications. Parallel STL160N10F8 MOSFET configuration handles large current, capable of driving up to 47 A (RMS) under natural convection conditions. Current sensing uses 3-shunt differential method and supports both absolute and incremental encoders as well as Hall sensors in both differential and single-ended modes. Communication is CAN with overcurrent, overvoltage, overheat protection and regenerative braking external resistor management functions. Table 2 summarizes the positioning of the three boards.

7. Utilization of Computational Headroom: Condition Monitoring and AI-Based Automatic Tuning

The approximately 40% computational headroom confirmed in section 5 has no value by itself. Value comes from what is built on that headroom. The first application is condition monitoring and predictive diagnostics. Using Motor Pilot record function in MC-SDK v6 to log operation data, inputting it to NanoEdge AI Studio to generate anomaly detection library, then integrating alongside motor control algorithm allows motor control and anomaly detection to be performed simultaneously on the same MCU. NanoEdge AI library is available free for use on STM32G4 development boards. However, according to technical Q&A, NanoEdge AI Studio includes time-series based models but operates only within the library range provided; directly defining model structure or porting custom architectures is impossible. This differs from STM32Cube.AI, which directly converts PyTorch or TensorFlow models.

The second application is AI-based PI automatic tuning. The background involves endemic PID tuning problems. P, I, D parameters are mutually dependent, integral saturation causes overshoot, pursuit of fast response risks oscillation, while pursuit of stability causes delay. Mechanical backlash and wear, load variations demand retuning, and even motor replacement requires retuning. Therefore, ST has explored an approach delegating this iterative process to deep reinforcement learning (DRL). An agent takes the action of adjusting controller gains, receives positive reward for fast and stable response, negative reward for overshoot or oscillation, and learns good gain combinations autonomously. One episode consists of 100 repetitions of step response, and in the case presented in the presentation, the entire process converged in approximately 30 minutes. Figure 7 shows PI value convergence and step response improvement as episodes progress. As a result, without relying on tuning experience, users can find optimal PID combinations by simply setting target metrics (overshoot, settling time, etc.), and even experts save time. However, according to technical Q&A, this function was implemented through STM32Cube.AI and has not yet been officially released, requiring separate technical inquiry.

8. Application Boundaries and Limitations

Precisely understanding the nature of this solution is the starting point for application. As the presentation and technical Q&A consistently clarify, the boards presented are reference designs, not finished products. The solution does not have fixed upper and lower bounds for output and torque; rather, customization to customer specifications is assumed. ST supports hardware and firmware design related to its solutions in the customization process, but application-level coding must be performed directly by customers. Cost effectiveness also cannot be determined definitively. The inherent advantage of dual drive is simultaneous control of two axes using a single MCU and power supply, increasing system integration and synchronizing between axes; depending on system configuration, costs can actually increase.

Technical boundaries are also clear. The limiting factor for shared current sensing is PWM duty cycle, managed through MCSDK sample timing correction and duty limitation. Current control period operates stably up to 40 kHz based on STSPIN32G4 and MCSDK FOC firmware, though while 100 kHz configuration is possible, practical use is difficult due to CPU load limitations. Sensorless FOC based on ST's HSO sensorless standard can provide sufficient torque at speeds below 10 Hz depending on motor, but sensorless is rarely used in robotics requiring precision control, so development direction focuses on loss minimization and thermal optimization. AI-based PI automatic tuning is still in pre-official release stage. Additionally, development of a GaN-based robot joint solution was mentioned in Q&A but specifications remain undisclosed.

Returning to the initial tension, the problem of fitting multiple axes into narrow space has been solved by choosing to share sensors, layout, and controller instead of replicating circuits per axis. Constraints created by sharing (PWM duty cycle, sacrifice of independent drive capability in parallel configuration) are managed through firmware correction and reconfiguration flexibility, resulting in a 4 x 4.5 cm dual FOC board with 40% computational headroom, and enabling predictive diagnostics and automatic tuning on top of that headroom. This, together with the conditions designers must review when transitioning from reference to production design, represents the conclusion this white paper aims to convey.