Humanoid actuators

What a humanoid actuator contains

Mentee Robotics lists self-made actuators with fully integrated on-board sensors and drivers under the heading of full vertical integration on its MenteeBot page. ^[10] EngineAI's specification for the EngineAI PM01 names a crossed roller bearing at the joint output and an internal-rotor permanent magnet synchronous motor. ^[7] Booster lists a dual encoder under the actuators of the Booster K1. ^[6]

Apptronik says that actuation is at the heart of humanoid performance and that Apollo 2 uses its patented actuator technology. ^[5]

The Booster K1 lists 6 degrees of freedom per leg, and AGIBOT lists 6 degrees of freedom for a single leg of the AGIBOT A2 Ultra and A2 Lite. ^{[6] [3]}

Reading torque figures

The LimX Luna page lists a maximum motor torque of 200 N·m and adds a footnote: the maximum differs among motors, and the value shown is that of the highest-torque motor. ^[9]

The product page of the Noetix Bumi gives a maximum torque of 70 N·m for the joint modules. ^[12] The developer documentation lists a torque per joint: 5 N·m for each arm joint, 15 N·m for thigh yaw, 30 N·m for the ankles, 27 N·m for the waist and 60 N·m for hip pitch, hip roll and knee. ^[13]

The AGIBOT A2 Ultra and A2 Lite pages give a peak knee torque of 270 N·m. ^[3] Booster gives a max peak torque of 60 N·m for the actuators of the Booster K1. ^[6]

EngineAI's highlights for the PM01 describe an integrated joint module delivering a peak torque of 130 N·m, while its specification table gives a max torque of 164 N·m and a peak torque density of 145 Nm/kg. ^[7]

When you compare robots, ask two things of every figure:

• Which joint does it belong to? LimX says its figure is that of the highest-torque motor. ^[9]
• Is it a peak, a maximum or a per-joint value? AGIBOT, Booster and EngineAI each use different wording. ^{[3] [6] [7]}

Transmission choices: cycloidal and tendon drives

The Berkeley Humanoid Lite authors built the actuators around a modular 3D-printed gearbox. ^[1] They note that 3D-printed gearboxes have reduced strength and durability compared with metal ones, and that they adopted a cycloidal gear design to address this. ^[1] They also report extensive testing of the 3D-printed actuators to validate durability, and a total hardware cost under $5,000. ^[1]

Agility says that Agility Digit 5 has a new leg design engineered to withstand repetitive lifting, powered by proprietary cycloidal actuator technology, and that it can repeatedly lift loads of up to 50 lb. (22.7 kg). ^[4]

Tendon drives take a different route. 1X describes the 1X NEO as using low inertia tendon drives, and says its Tendon Drive actuation creates precise, low-energy movements necessary for home use. ^[11] The page lists a torque accuracy of 2% and a backdrivability of 95% for the Tendon Drive. ^[11] It also ties the tendon-driven actuators to gentle, safe movements, alongside a soft body wrapped in a lattice polymer. ^[11]

Agility ties the cycloidal actuators of Digit 5 to repeated lifting of loads. ^[4] 1X ties the tendon drives of NEO to gentle, safe movements in the home. ^[11] Whether either design suits your application is something to test yourself.

Efficiency, integration and supply

Apptronik says its actuator platform is designed for 90%+ energy efficiency, simple maintainability, mass manufacturability and supply chain resiliency, and that it reduces dependence on single-source suppliers. ^[5]

For Figure 03, Figure says it chose to integrate vertically across many critical modules, including actuators, batteries, sensors, structures and electronics, all designed completely in-house. ^[8] Figure also says the actuators perform at 2x faster speeds with improved torque density, which it ties to faster pick and place. ^[8]

Mentee Robotics also lists full vertical integration, with self-made actuators, bespoke electronics and communication. ^[10]

Interfaces, control and protection in practice

The documentation of the Noetix Bumi lists the joint communication method as EtherCAT and CAN FD. ^[13] For application development it describes HighController, which builds higher-level scheduling on the manufacturer's motion control strategy without changing the underlying control logic. ^[13] For motion control development it describes LowController, which opens the motor control interface of the motion control program to the SDK and lets users adjust low-level motor parameters directly at a control frequency of 500Hz. ^[13] The documentation says that only the EDU version of the Bumi can be used for LowController development. ^[13]

Each motor command in LowController carries position, velocity, torque, kp and kd values, and the documentation tells developers to follow the parameter ranges it lists per joint. ^[13] Motor state comes back with position, velocity, torque, temperature and an error code. ^[13] The error codes include overcurrent, low voltage, encoder error, MOS over-temperature, motor coil over-temperature, communication loss and overload. ^[13]

The documentation describes a protection mode: when the posture leaves the stable standing range beyond a preset threshold, the whole-body joints are disabled and motion control stops. ^[13] It also warns against lifting the robot off the ground in the whole-body walking gait, because the controller may keep outputting joint torque and put abnormal load on the joints and transmission parts. ^[13]

The Berkeley Humanoid Lite authors report a locomotion controller trained with reinforcement learning and transferred zero-shot from simulation to hardware. ^[1] The authors of the original Berkeley Humanoid describe a narrow sim-to-real gap and a simple reinforcement learning controller with light domain randomisation on their lightweight, in-house-built robot. ^[2]