Engine AI SE01 solved the gait problem that bipedal humanoid robots have struggled with for 20 years. Most bipedal robots use model-based locomotion controllers, engineers specify gait parameters (step length, step frequency, center-of-mass trajectory) and a controller executes them while maintaining balance. The result: stable but mechanically stiff walking with short strides, no arm counterswing, and the characteristic robot shuffle. Engine AI replaced this with an end-to-end neural network trained on human motion capture data and fine-tuned through reinforcement learning in simulation. The network learned natural gait emergently, long strides, heel-toe contact, hip rotation, arm counterswing, because it was trained to reproduce human motion, not execute an engineer-specified pattern.
Technical Specifications
- Height: 170 cm
- Weight: 55 kg
- Total DOF: 32
- Leg DOF: 6 per leg = 12 total (hip 3, knee 1, ankle 2)
- Arm DOF: 4 per arm = 8 total
- Hand DOF: 6 per hand = 12 total (5-finger)
- Hip joint torque: 330 Nm (self-developed harmonic drive actuators)
- Walking speed: 2 m/s maximum, highest of any Chinese humanoid at launch
- Capabilities: walk, squat, grasp, run, jump
- Battery: 2-3 hours continuous
- Compute: dual-processor, NVIDIA GPU (AI inference) + Intel CPU (real-time control)
- Vision: 6 HD cameras (360-degree coverage) + Intel RealSense D435 depth camera
- LiDAR: 360-degree 3D environment mapping
- Body material: aerospace-grade aluminum alloy
- Price: CNY 150,000-200,000 ($20,600-$27,500)
Neural Gait Controller Architecture
- Input: 77-dimensional vector at 500 Hz, joint positions (32), joint velocities (32), IMU data (6), foot contact signals (4), velocity command (3)
- Architecture: 6-layer transformer, 256 hidden dimensions
- Training method: reinforcement learning in MuJoCo simulation with domain randomization (floor friction, foot mass, joint damping varied randomly per episode)
- Pre-training data: CMU MoCap + AMASS human motion capture datasets
- Output: 32 joint position targets at 500 Hz fed to low-level PD controllers
- Real-world transfer: no additional hardware training required, simulation-trained policy runs directly on SE01
The 330 Nm hip joints are structurally critical to the neural gait result. Natural human walking requires powerful hip extension at push-off and controlled hip flexion at swing initiation. Robots with underpowered hip actuators compensate by shortening stride, producing the robot shuffle. SE01s 330 Nm hip torque eliminates this constraint, allowing the neural network to produce the full-extension, long-stride gait that human motion capture data contains.
6-Camera Perception System
- 2 forward cameras: stereo pair for depth estimation and object detection
- 2 lateral cameras: left and right peripheral awareness
- 2 rear cameras: obstacle detection and workspace awareness
- Processing: NVIDIA GPU processes all 6 streams simultaneously for real-time 3D world model
- Combined with 360-degree LiDAR: full environment map for navigation and obstacle avoidance
PM01 Wheeled Variant
The PM01 uses the SE01 upper body (arms, hands, compute, all sensors) on a wheeled omnidirectional base. Battery life: 8 hours vs 2-3 hours bipedal. Zero fall risk. Higher payload stability. Identical software stack, task AI developed on PM01 transfers directly to SE01. Engine AI customers can pilot PM01 on flat floors and deploy SE01 for stairs or outdoor environments without software migration.
Commercial Status
- Launch: October 2024, NVIDIA and SenseTime engineers assessed gait as technically novel
- Units delivered 2025: 150-500
- Pricing: $20,600-$27,500, most affordable full-capability industrial humanoid in China
SE01 neural gait close look: Engine AI SE01 article. PM01 wheeled: Engine AI PM01.
