Locomotion & Balance: Teaching Robots to Walk

Bipedal Walking Technology

Walking on two legs requires continuous, real-time coordination of dozens of joints and muscles — maintaining a moving center of mass over a small, shifting support polygon while adapting to terrain variation, unexpected disturbances, and changing task demands. For robots, this is achieved through a combination of mechanical design (stiff or compliant joints, foot geometry), onboard sensing (IMUs, foot pressure sensors, joint encoders), and control software that runs hundreds of times per second to keep the robot upright and moving.

Two dominant control paradigms have emerged. Model predictive control (MPC) plans future steps based on a physics model of the robot — predicting where the center of mass will be 0.5–2 seconds ahead and computing joint commands to maintain stability. MPC is deterministic, interpretable, and well-suited to structured environments, but struggles with unexpected disturbances. Reinforcement learning (RL) takes a fundamentally different approach: learning walking policies through millions of simulated trials and then transferring them to the real robot. RL policies tend to be more robust to terrain variation and unexpected pushes, but are less predictable and harder to debug. Leading Chinese companies — particularly Unitree, which released China's first running humanoid (H1) in 2023 — increasingly combine both approaches: MPC for nominal walking and RL for disturbance recovery and dynamic motions. The hardware requirements for dynamic locomotion are demanding — see our Actuators & Motors report for the joint specifications needed.

Gait Types and Terrain Adaptation

Walking, trotting, running, and stair climbing each require different gait patterns and joint coordination strategies. Current Chinese humanoid robots implement several gait modes: a slow, stable walk for flat terrain (0.5–1.0 m/s); a faster dynamic walk for open spaces (1.0–2.0 m/s); stair climbing gaits with shorter step lengths and higher knee flexion; and running gaits with aerial phases. Terrain sensing — combining foot pressure sensors, lidar, and depth cameras — enables real-time adaptation to slopes, steps, and obstacles. The leading hardware configuration for robust terrain adaptation pairs compliant ankle joints (allowing the foot to conform to uneven surfaces) with high-bandwidth knee and hip actuators that can rapidly adjust stance height.

Joint Requirements for Dynamic Locomotion

Dynamic walking and running place extreme demands on joint actuators: they must switch rapidly between high-torque (stance phase, supporting 50–100% of body weight) and high-speed (swing phase, moving the leg forward in under 200ms) operation, recover from unexpected impacts, and do all of this with minimal energy consumption. The key specifications are torque density (Nm/kg — higher allows lighter robots), bandwidth (how fast the motor can change its output — 50Hz+ needed for dynamic balance), and backdrivability (how easily external forces can move the joint — critical for safe human interaction and fall recovery). Quasi-direct-drive actuators — high-torque motors with low-ratio planetary reducers — have become popular in research-grade platforms for their excellent backdrivability. Production robots often favor harmonic reducer-based joints that sacrifice some backdrivability for higher torque density and precision.

Dynamic Balance

Balance control is what separates a humanoid robot from a statue. When a robot is pushed, bumped, or steps on an unexpected surface, its balance controller must instantly compute compensatory joint commands to prevent a fall — all within milliseconds. This requires not just fast computation but a detailed, continuously updated model of the robot's state (center of mass position and velocity, foot contact forces, joint positions) and a control law that can generate stable recovery motions from any starting condition.

Three main balance control techniques are used in modern humanoid robots. Zero-moment point (ZMP) control — the classical approach — ensures the robot's ground reaction force stays within the support polygon defined by its feet. ZMP is reliable on flat terrain but struggles with dynamic motions where the robot briefly leaves the ground. Capture point methods compute where the robot must step to avoid falling, enabling more aggressive recovery strategies. Whole-body control (WBC) frameworks simultaneously satisfy balance, task execution, and joint limit constraints through real-time optimization — the most capable but also most computationally expensive approach. Chinese companies increasingly combine WBC with reinforcement learning-trained recovery policies, where the RL handles unexpected disturbances that exceed the WBC's model-based predictions.

Foot Design and Ground Contact

The foot is the robot's only contact with the ground, and its design has enormous implications for balance and locomotion performance. Flat rigid feet are simple but offer limited stability on uneven terrain. Compliant feet with embedded force sensors can measure ground reaction forces in real time, enabling far more sensitive balance control. Some designs include articulated toes for improved push-off efficiency in walking. Among Chinese humanoid platforms, foot design philosophies vary: Unitree favors compact, sensor-rich feet optimized for dynamic running; UBTECH's Walker series uses larger, flatter feet for stability during industrial tasks; and several research platforms are experimenting with multi-point articulated feet inspired by human biomechanics.

Running & Acrobatics

Running introduces a new challenge not present in walking: flight phases, where both feet are simultaneously off the ground. During these phases, the robot cannot generate ground reaction forces and must rely entirely on its pre-flight state and in-air joint configuration to land safely. Running requires significantly more actuator bandwidth, faster balance recovery, and much higher structural loads than walking — typically 3–5x the ground reaction force of normal walking.

Unitree's track record in running humanoids is the benchmark: the H1 can sustain continuous running, and the H2 demonstrated highly dynamic choreographed martial arts at the 2026 Spring Festival Gala. But perhaps the most technically impressive dynamic motion demonstration came from MagicBot's Z1, which performed the "Thomas 360° breakdance flare" at the same event — requiring the robot to support its full body weight on its hands while rotating its legs in a 360° horizontal circle, generating extreme centripetal forces that demand exceptional hip mobility, precise center-of-mass control, and impact-resistant joint design.

Spring Gala Demonstrations

The CCTV Spring Festival Gala — China's most-watched annual television event — has become an unofficial showcase for China's humanoid robot capabilities. In 2025, 16 Unitree H1 robots performed an all-AI-driven collective dance (《秧BOT》) — the first time humanoid robots autonomously choreographed formation changes on live television. In 2026, the Gala featured an even more ambitious demonstration: 25 Unitree robots performing the martial arts routine 《武BOT》, the MagicBot Z1 executing a world-first 360° Thomas flare, Galbot's G1 demonstrating dexterous hand tasks, and Songyan Power's E1/N2 engaging in real-time AI conversation with comedian Cai Ming.

The 2026 Spring Festival Gala demonstrations revealed both genuine capability and careful choreography. According to Unitree's IPO prospectus, the 《武BOT》performance required upgrading their cluster planning control system and stunt motion control technology, developing a high-dynamic formation running control system and large-scale pre-trained motion control model. The prospectus specifically credits innovations in AI full-body state modeling and elastic energy management systems for solving localization challenges during violent motion and handling transient high-current impacts. Multiple world records were set: first continuous parkour, first catapult backflip exceeding 3 meters, first single-foot consecutive flips, first Airflare 7.5 rotations, and fastest autonomous formation running at 4m/s. On the other hand, these were choreographed performances on controlled surfaces — different from autonomous operation in unstructured factory environments. The commercial impact was significant regardless: participating companies reported surges in investor interest and customer inquiries in the weeks following the broadcast.

• Unitree (25 robots): Demonstrated multi-robot synchronization with multiple world-record stunts. Also won 11 medals (most golds and most total) at the inaugural World Humanoid Robot Games (Aug 2025) — including 1st place in 1,500m (6:34), 400m (1:28), 4×100m relay (1:48), and 100m obstacle course (33.71s)
• MagicBot Z1: Thomas 360° flare requires extreme hip mobility, precise center-of-mass control, and impact-resistant joint design
• Galbot G1: Dexterous manipulation tasks (clothes folding) demonstrated fine motor skill in a dynamic environment
• Songyan Power E1/N2: Real-time AI conversation with live performers demonstrated natural language integration in a live, unscripted setting