Moving the inverter into the joint is not a design philosophy. It is a consequence of removing the heat sink, and the heat sink goes away only when the power stage is efficient enough to do without it. TI puts the enabling condition where it belongs: GaN with the driver integrated can push power stage efficiency beyond 99 percent, which allows integrated motors to reduce or eliminate the heat sink [1]. Teams that try to decentralise without first winning the thermal argument end up putting a heat sink inside a joint, which is worse than leaving the drive in the cabinet.
Understanding the causal order matters, because it tells you which problem to solve first.
Two architectures, one difference
The classical industrial robot is centralised. Every servo drive control module lives in the controller cabinet, and each axis is reached by a cable carrying the phase lines, the brake and the position sensor. A manipulator typically carries three to seven axes [2], so the cabinet holds three to seven drives and the harness carries all of them.
The decentralised version moves each control module out to its axis [2]. The cabinet keeps the compute, the communication module and the power supply. The joint gets the inverter.
The immediate consequence is the harness. Centralised wiring runs more than ten wires per motor. Decentralised wiring can put a whole robot arm on a two wire communication bus [1], with DC power distributed alongside it. That matters because a manipulator cable tree is typically replaced every two to three years on the assumption of ageing, without anyone testing whether it failed [3]. Fewer conductors is less a cost saving than a service life unlock.
The chain that makes it possible
Four things had to arrive together.
Efficiency beyond 99 percent in the power stage. The enabling condition, reached by GaN with the gate driver integrated [1]. Below that level, in our experience, the dissipation inside a sealed joint needs a thermal path you do not have.
PWM resolution measured in picoseconds. TI cites picosecond PWM resolution [1]; in our reading that is what lets high switching frequency deliver current resolution rather than quantisation noise, and what makes a compact joint drive controllable rather than merely small.
Current sensing that survives next to the switching node. Current sensors with a ±50 µV offset and the power supply integrated into the same package deliver both the loop accuracy and the board area reduction the joint needs [1]. Precision and miniaturisation turn out to be the same requirement, not competing ones.
A network that fits inside the joint. TI names single pair Ethernet as the network for the joint [1]. In our reading the physical layer transceiver is what turns the two wire bus from a diagram into a product.
Remove any one and, in our experience, the architecture does not close. That is why decentralisation arrived as a cluster of capabilities rather than a gradual migration.
What the standards define, and why it matters for scope
TI's overview, following ISO 8373:2012, partitions a robot system into five blocks: the controller system, the manipulator, the teaching pendant, vision and sensors, and the end effector [2]. The definitions are worth knowing precisely, because they draw the boundary of what a joint module supplier is actually selling.
The controller system is defined as the set of logic control and power functions that allow monitoring and control of the mechanical structure and communication with the environment. The manipulator is the jointed mechanism itself, and the definition explicitly excludes the end effector [2].
Decentralisation moves power functions out of the controller system block and into the manipulator block. That is a redistribution of responsibility across a standards boundary, which is why the safety case has to be revisited rather than inherited. ISO 10218 governs robot safety requirements, and IEC 61800-5-2 governs drive safety subfunctions such as safe torque off [1]. When the drive physically relocates, the argument for how those functions are achieved relocates with it.
The DC bus range tells you which market you are in
Industrial servo drives span DC bus voltages from 48 V to 1200 V [2]. That single range explains most of the divergence in the design literature.
From 300 V to 1200 V DC link you are dealing with IEC 61800-5-1 and basic or reinforced insulation as a requirement [2], and, at the top of that range, with IGBTs whose switching frequencies sit in the low to medium band of 5 kHz to 20 kHz [2]. For 12 V to 60 V fed inverters, non isolated current sense amplifiers become attractive on system cost [2]. In our view that turns isolation into a noise management choice rather than a safety requirement. GaN stages have been tested at 40 kHz, 60 kHz and 80 kHz with almost identical board losses at 40 kHz and 80 kHz [4].
In our reading, humanoid and collaborative robotics sits firmly at the bottom of that voltage range. Design guidance written for the top of it transfers in method but not in component selection, and treating the two as one market is how teams over specify their way out of the form factor.
Do not forget the other motors
One distinction that gets lost in the focus on joints. A robot contains two populations of motors.
The high performance population is the joint actuators: field oriented control, GaN power stages, absolute encoders, high resolution current sensing, everything discussed above.
The other population is auxiliary. Fans and pumps for cooling and hydraulics [1]. These want a single chip motor controller with the power stage integrated, six MOSFETs plus a digital core running trapezoidal control in the part TI names [1]. They do not need FOC, encoders, or a picosecond PWM engine, and giving them those things wastes cost, board area and design effort.
Both populations appear in every robot. Sizing the second one with the first one's architecture is a common and expensive mistake.
The question decentralisation leaves open
If every joint now has an inverter, a controller, a sensor chain and a network interface, the robot has become a distributed real time system, and distributed real time systems fail differently from centralised ones.
Clock synchronisation across joints, deterministic delivery of setpoints, coordinated safe states across independent controllers: none of these were problems when all the drives lived in one cabinet on one backplane. They are problems now. The thermal argument made decentralisation possible, but the remaining work is mostly in timing and coordination rather than in power electronics.
References
- Texas Instruments, How Motor Drive Innovations Are Helping Solve Robotic Movement Design Challenges, SSZTCZ3, September 2023.
- Texas Instruments, An Engineer's Guide to Industrial Robot Designs, SSIY006, 2Q 2020, chapter 1 and sections 2.2.1 and 2.2.11.
- Texas Instruments, How and Why to Use Single Twisted-Pair Ethernet (SPE) for Industrial Robotics Environments, SNLA420A, September 2022, revised January 2024.
- Texas Instruments, Application of GaN FET in Humanoid Robots, SNOAAB3, January 2025.