Engine AI SE01: How 330 Nm Joints and Neural Gait Redefined Chinese Humanoid Walking

When Engine AI unveiled the SE01 in October 2024, the robot did something none of its Chinese competitors had demonstrated convincingly: it walked with a gait that looked genuinely human. Not the stiff, short-stride walking common to bipedal robots balancing against falling, a smooth, long-stride, heel-toe gait with natural hip rotation and arm swing. The key was not a new mechanical design. It was an end-to-end artificial neural network trained to optimize gait performance directly from sensor data, bypassing the traditional model-based locomotion controllers that produce the characteristic robot shuffle. This article examines the SE01’s engineering and its commercial positioning in China’s humanoid market.

The Gait Problem in Humanoid Robots

Most humanoid robots walk using model-based controllers: engineers specify a gait pattern (step length, step frequency, center-of-mass height trajectory) and the controller executes it while maintaining balance through a separate stability computation. The limitation of this approach is that the gait is constrained by the model’s accuracy. Real human walking involves hundreds of micro-adjustments per step, hip tilt, ankle push-off timing, arm swing counterbalance, that are difficult to specify explicitly but are essential for the fluid appearance and energy efficiency of natural gait. Model-based controllers produce stable but mechanically-looking walking because they optimize for stability rather than naturalness.

Engine AI’s approach is different: train an end-to-end neural network on motion capture data of human walking, then fine-tune it on the SE01’s specific kinematic chain through reinforcement learning in simulation. The network takes sensor inputs (joint states, IMU data, foot contact signals) and outputs joint torque commands directly, without an intermediate gait model. The resulting gait emergently reproduces human-like features, long strides, heel-toe contact, natural arm counterswing, because the network learned them from human data rather than being programmed with them.

SE01 Specifications

Height 170 cm, weight 55 kg. Degrees of freedom: 32 total, including 12 hand DOF (6 per hand) and 8 arm DOF (4 per arm). Walking speed: 2 m/s, faster than any competing Chinese humanoid at launch. The SE01 can squat, grasp, run, and jump. Actuators: self-developed harmonic joint modules with maximum 330 Nm per joint, a torque density that enables the powerful hip and knee movements that make natural gait possible. Higher peak joint torque allows longer strides without the characteristic short-step constraint of underpowered bipedal robots.

Compute: dual-processor architecture with NVIDIA GPU for AI inference and Intel CPU for sensor processing and real-time control. Sensor suite: Intel RealSense D435 depth camera, 360-degree LiDAR, six HD cameras for 3D environment mapping and obstacle avoidance, plus proprioceptive sensors at all joints. Aerospace-grade aluminum alloy body. Price: CNY 150,000-200,000 ($20,600-$27,500) at launch, positioning it between Unitree’s G1 ($16,000) and UBTECH’s Walker S2 (enterprise pricing).

NVIDIA and SenseTime Validation

Engine AI’s SE01 attracted attention from NVIDIA and SenseTime immediately after its October 2024 unveiling. Digitimes reported that engineers from both companies were present at the SE01 demonstration and assessed the gait performance as genuinely novel, the neural network approach producing smoother and more energy-efficient walking than any alternative implementation they had evaluated. NVIDIA’s interest is directly commercial: Engine AI’s dual-processor architecture uses NVIDIA GPU for inference, making Engine AI a natural NVIDIA robotics platform partner. SenseTime’s interest is in computer vision integration, the SE01’s 6-camera sensor suite creates an opportunity for SenseTime’s vision AI stack to power SE01 perception.

Commercial Status in 2026

Engine AI delivered 150-500 units in 2025, placing it in the second tier of Chinese humanoid commercialization behind AgiBot and Unitree but ahead of most domestic competitors. The company has not disclosed order backlog or 2026 production targets publicly. The SE01’s price point ($20,600-$27,500) positions it as the most affordable full-capability humanoid with industrial ambitions in China, below UBTECH Walker S2 but above Unitree G1 in feature richness for industrial applications. The 2 m/s walking speed and 330 Nm joints make it the physically capable choice for applications requiring fast transit between workstations and higher payload manipulation.

PM01: The Wheeled Variant

Engine AI also produces the PM01, a wheeled humanoid on the same upper body platform as the SE01, replacing bipedal legs with a wheeled mobile base. The PM01 trades the SE01’s walking capability for the operational simplicity, stability, and 24/7 runtime of wheeled platforms. Engine AI positions the PM01 for fixed factory floor applications where bipedal mobility is not required, and the SE01 for environments with stairs, ramps, or bipedal-specific constraints. This dual-platform strategy mirrors AgiBot’s A2/G2 split and Unitree’s G1/G1-D architecture.

For the full China humanoid competitive field: Top 40 China Robot Rankings 2025. Technical comparison with competing platforms: Fourier GR-2 vs Rokae HumanX force control.

Sources

  • Digitimes: China humanoid robot stuns NVIDIA and SenseTime with human-like walk (January 2025)
  • Maginative: Engine AI unveils SE01, impressive robot that walks like a human (October 2024)
  • Aparobot / Robots International: EngineAI SE01 full specifications
  • Engine AI: Product documentation and pricing (2025)

Related posts

Leave the first comment