BXI Elf3

A highly dynamic humanoid with 31 DoF, standing 145 cm tall. Run, jump, dance — field-proven motion performance for research and development.

Exterior Design

A refined design combining human-like proportions with aluminum and carbon components. With 31 DoF across the whole body, it naturally expresses a wide range of postures — standing, crouching and sitting.

Highly dynamic motion, proven in the real world.

Motion control validated not only in simulation but in real competitions and field use. The robot reliably performs highly dynamic movements such as running, jumping, dancing and climbing over obstacles.

Joint Modules Built for Humanoids

The BXI-50/70/85 hollow planetary series. Just three motor types are enough to build a humanoid with more than 30 DoF, greatly reducing design, manufacturing and maintenance costs.

Modular designThree motor types cover the whole body, with a torque range of 25–150 N·m that can be tuned and customized.

Planetary + hollow structureThe hollow shaft routes cables internally for better reliability and range of motion. The planetary gearbox is highly shock-resistant and withstands falls.

True dual encodersDirect measurement on both the motor and output sides improves end-effector control accuracy.

Field-provenValidated through demanding tests and competitions on multiple humanoids, and continuously improved.

Item	BXI-50	BXI-70	BXI-85
Rated torque	7 N·m	15 N·m	40 N·m
Peak torque	25 N·m	50 N·m	150 N·m
Peak phase current	30 A(rms)	60 A(rms)	90 A(rms)
Outer diameter × height	64 × 65.5 mm	81 × 66.5 mm	99 × 71.5 mm
Hollow diameter	6 mm	9 mm	10 mm
Weight	0.5 kg	0.8 kg	1.4 kg
Common specifications	Rated voltage 24–48 V / no-load speed 200 rpm / rated speed 100 rpm / gear ratio 19.5 / cross-roller bearing / dual encoders (motor side: magnetic, output side: inductive) / MIT protocol with hybrid torque, position and velocity control / CAN, CAN FD (RS485 coming soon)

Motion Control ("Cerebellum") Solution

Reinforcement-learning-based motion control enables highly dynamic whole-body movement. With a systematized pipeline from data to the real robot, it is ready for exhibitions and performances.

Automatic conversion of human motion into robot joint trajectories

Reinforcement learning in parallel simulation environments

Quality validation in high-fidelity simulation

Deploying the control policy to real hardware

Learn in simulation, deploy on the real robotIn a parallel simulation environment, we train the control policy with reinforcement learning so that BXI Elf3 tracks a reference motion (green) generated from human motion data. Even complex whole-body movements, including falling down and getting back up, are validated in Sim2Sim and then deployed on the real robot.

Reinforcement learning
Whole-body motion tracking
Sim2Sim validation
Real-robot deployment

Main Applications

Exhibitions & performancesDemonstrations of dynamic motion such as walking, running, jumping and dancing

Guidance & patrolGuiding visitors in exhibition halls and facilities with precise navigation and multilingual AI commentary

Teleoperation & data collectionCollecting AI training data with VR, exoskeletons and data-capture devices

Research & developmentA platform for embodied AI and VLA model research, or for developing your own robot products

Combine with Hands and Bases

Combine HumansX Physical AI solutions to configure a robot for your application.

Product page ↗Request materials