TL;DR
The NING Humanoid is a new 1.1-meter, 20kg robotic platform designed through "concurrent engineering"—where hardware and control algorithms are developed in tandem. By optimizing the robot’s physical mass distribution to match its mathematical models, the researchers achieved a highly agile system capable of stair climbing, push recovery, and simulated backflips using low-cost, high-torque actuators.
What Problem Does This Paper Address?
In the world of humanoid robotics, there is often a "performance gap" between a robot’s physical capabilities and the mathematical models used to control it. High-end robots like Boston Dynamics' Atlas use complex hydraulic systems that are powerful but expensive and difficult to maintain. On the other end, many affordable humanoids use Quasi-Direct-Drive (QDD) motors but struggle with agility because their mass is distributed in a way that makes the underlying control math (like centroidal dynamics) inaccurate.
The researchers behind the NING Humanoid set out to build a platform that is:
- Agile and Robust: Capable of dynamic tasks like jumping and flipping.
- Low Cost: Utilizing modular QDD actuators rather than expensive hydraulics or series-elastic actuators.
- Mathematically Optimized: Designed so that its physical structure aligns with simplified control models, reducing the "approximation error" that often leads to instability.
Methods
The team employed a concurrent design approach, focusing on three main pillars:
- Hardware Co-design (CII): They utilized the concept of Centroidal Inertia Isotropy (CII). By tilting the hip roll and yaw actuators at 45-degree angles, they created a "compact crotch" design. This concentrates the heavy motor mass near the robot's Center of Mass (CoM) and reduces the weight of the limbs, making the leg's inertia more predictable for the controller.
- Transmission Mechanism: To ensure high rigidity and linear torque transmission, the team used a four-bar linkage mechanism for the knee and ankle joints. This avoids the elasticity issues found in belt-driven systems while remaining simpler to control than complex multi-bar linkages.
- Control Framework (WB-MPC): The robot uses a Whole-Body Model Predictive Control (WB-MPC) framework.
- The MPC layer predicts future states and plans step placement using a simplified centroidal model.
- The WBC (Whole-Body Control) layer solves a convex Quadratic Programming (QP) problem at 1kHz to optimize joint torques and accelerations in real-time, ensuring the robot follows the planned trajectory while maintaining balance.
Datasets
As this is a hardware and control systems paper, it does not use traditional "training datasets." Instead, the system was evaluated using:
- MuJoCo Simulation: A 3D URDF (Unified Robot Description Format) model was used to simulate mass, inertia, and dynamic tasks like back-flipping.
- Real-World Experimental Scenarios: The robot was tested on diverse terrains including rough grass, snow, 10-degree slopes, and low stairs to validate its push recovery and locomotion stability.
Key Insights and Ideas
- Design for the Math: The most significant "aha moment" is the intentional alignment of hardware with the Centroidal Dynamic Model. By ensuring the leg's CoM stays close to the body's CoM regardless of posture, the simplified math used in the MPC becomes a much more powerful tool for real-world balance.
- The 45-Degree Hip Tilt: This specific mechanical choice solved two problems at once: it made the robot more compact and significantly lowered the inertia of the legs, which is critical for high-speed, agile movements.
- High-Bandwidth Communication: Operating at a 1kHz control bandwidth via EtherCAT allows the robot to react to "unknown disturbances" (like a shove or a slippery patch of ice) almost instantaneously.
Key Results
- Physical Specs: 1.1 meters tall, 20 kg weight, and 18 Degrees of Freedom (5 per leg, 4 per arm).
- Agility: Successfully demonstrated stable walking on grass and snow, as well as climbing stairs and 10-degree inclines.
- Robustness: Demonstrated "push recovery," where the robot maintains balance after being physically disturbed.
- Simulation Success: The WB-MPC framework successfully executed a backflip in the MuJoCo simulation environment, showcasing the potential for extreme acrobatics.
- Efficiency: The use of QDD actuators provided smooth back-drivability and accurate torque feedback without the high cost of specialized sensors.
Limitations and Future Work
While the NING Humanoid is highly capable, the authors note a few areas for growth:
- Sim-to-Real Gap for Acrobatics: While the backflip was successful in simulation, it has not yet been demonstrated in the physical world.
- Perception: The current version focuses heavily on locomotion and balance. The authors plan to "augment perception-based state estimation," essentially giving the robot better "eyes" (vision/LiDAR) to plan its path through complex environments autonomously.
- Trajectory Planning: Future work will involve more advanced trajectory planning to allow the robot to perform even more complex, non-periodic movements.
Summary updated