EngineAI T800 Humanoid Robot Technical Analysis

Start with the configuration

EngineAI T800 is a family of full-size electric humanoids. The Development Edition manual describes a 1.73 m robot weighing approximately 84.5 kg with its battery and having 25 joint degrees of freedom. Pro and Max add hardware. A specification copied from one version can give the wrong answer about another version's arms, hands or sensors. [1]

EngineAI was established in Shenzhen in October 2023. Its official team page identifies Zhao Tongyang as founder and CEO. The company's T800 launch notice is dated 8 December 2025. This analysis compares its public product pages with development documentation updated in August 2026, checked again on 8 October 2026. [2] [3] [4]

Basic, Development, Pro and Max

The following table follows the English product page. Development Edition is called Open Source Edition there. These are manufacturer listings, with discrepancies addressed below. [5]

[5]

Degrees of freedom count independent joint movements. They do not measure how many objects a robot can grasp. Extra wrist motion changes reachable hand orientations; finger joints change possible contacts. Task performance still depends on control, sensing and the object being handled.

Where the official pages disagree

[6] [5] [1] [7] [8] [9]

Motors move different loads

T800's commercial specification lists aluminum and titanium construction with active air cooling. The joint manual describes Q-series actuators using planetary reductions and dual encoders. Its motor table ranges from 50 N·m for Q25H to 415 N·m for Q300H-L. The paragraph before that table gives Q25H as 52 N·m, an unresolved discrepancy within the manual. The product's 450 N·m headline therefore cannot be assigned to every joint, and its difference from the detailed motor table needs clarification. [6] [10]

Torque is the turning effect of a force about a joint. A hip supporting the torso has a different load from a wrist orienting an empty hand. Peak torque also differs from the torque a motor can hold throughout a work cycle. Repeated current through a motor creates heat. Cooling removes some of that heat, but a cooling description alone does not establish a continuous payload rating.

From sensor measurements to joint state

The Development sensor manual lists an Odin2 camera, head-mounted LiDAR and an inertial measurement unit in the torso. The IMU reports acceleration and angular motion. Its quoted 0.05° roll and pitch accuracy applies to stated static conditions at 25°C without prolonged acceleration. That number is not a measurement of the robot's balance accuracy during a kick. [9]

The dedicated computer document separates a Nezha Robot PC from a Jetson AGX Orin AI PC with 64 GB RAM. Both run Ubuntu. The robot's electrical interface includes Gigabit Ethernet. These details establish computing and communication hardware; they do not identify the perception network running in a particular video. [8] [1]

What the developer can command

EngineAI's ROS 2 interface accepts desired joint position, velocity, stiffness, damping and feed-forward torque. Measured joint position, velocity and torque return through a separate state topic. The documented controller combines position error, velocity error and feed-forward torque. Its guidance recommends a command frequency of 100–500 Hz. This describes the interface, not a measured policy inference rate. [11]

The manual supplies separate examples for holding a standing posture and running a reinforcement-learning policy. A fixed target supplies the former; a network using sensor observations supplies the latter. The motion-state table also distinguishes standing, walking, lower-body balance and recovery movements. A controller available on PM01 must not be assumed available on T800 merely because both appear in the same manual. [11] [12]

The relationship between estimated body motion, foot contact and motor commands is explained in how humanoids maintain balance while walking.

Walking figures and fighting footage

The Development manual lists forward walking at 0–1.5 m/s. Its separate running entry is below 3 m/s and is labelled as hardware-supported capability. The commercial page's speed entry is at least 2 m/s with a similar qualification. Neither entry supplies a measured run with distance, duration, surface and success count. [1] [6]

EngineAI labels T800 sequences as punches, kicks and walking. Those labels alone do not establish autonomous opponent tracking. URKL's official resources provide 30 Hz human motion data and retargeted T800 joint sequences alongside simulation materials. A training resource shows one available development route; it does not identify the controller used in every physical performance. [5] [13]

A strike moves the torso and changes the load on the supporting foot. Read the control and balance constraints in humanoid fighting for the distinction between a stored motion and a sensor-driven response.

What a purchase quote must resolve

EngineAI's notice dated 8 December 2025 announced T800 from RMB 180,000. It did not give a complete purchase configuration or delivery terms. A current quote should identify the edition, hands, computers, LiDAR, batteries and support. No independent endurance test or repeated industrial task success rate was identified in the sources reviewed. [4]