Technical Rules and Engineering Challenges of Humanoid Robot Half-Marathons

The 2026 Beijing Yizhuang Humanoid Robot Half-Marathon is scheduled for April 19. Unlike traditional demonstration-based robotics events, this competition places humanoid robots in continuous motion tasks within an outdoor environment that closely resembles real-world conditions. Through explicit rules regarding timing, robot swapping, battery swapping, collisions, and cutoff times, it tests the stability and autonomy of robotic systems over long-distance operations.

From an engineering perspective, completing a half-marathon is not simply about “making the robot walk faster.” It requires simultaneously solving path planning, dynamic balancing, visual perception, energy management, fault recovery, human-robot collaboration, and safety control. The value of the competition rules lies in transforming these inherently scattered technical capabilities into quantifiable, comparable, and reviewable comprehensive metrics.

1. Technical Goals Reflected in Competition Rules

The total race distance is 21.0975 kilometers. The route includes flat ground, slopes, curves, and narrow sections, encompassing over 10 types of terrains and multiple left and right turns, with some curves approaching 90 degrees. This means robots cannot rely on a single straight-line walking strategy; they must switch motion states under continuously changing road conditions.

From a control system perspective, the following aspects require attention:

1. Gait Stability
   When walking on flat ground, the support area is relatively stable. Upon entering a slope, the center of mass, foot landing points, and joint forces change. If fixed stride lengths and joint parameters are maintained, issues such as center of mass deviation, foot slipping, or insufficient landing are likely to occur.

2. Turning Ability
   Right-angle curves require the robot to identify the road direction in advance and adjust its speed and gait based on curvature. Turning too early may lead into non-track areas, while turning too late may cause sudden stops, collisions, or falls. Therefore, path planning must consider the turning radius, body orientation, and foot reach, not just the target point.

3. Narrow Section Navigation
   Narrow areas compress the walkable space and increase the probability of contact with guardrails, other equipment, or staff. The system needs to dynamically adjust walking speed and body posture based on environmental width, reducing stride length and increasing observation frequency if necessary, while prioritizing safety boundaries.

4. Long-Term Continuous Operation
   Short-term demonstrations can often rely on manual monitoring and temporary parameter tuning, but long-distance racing requires the system to maintain consistent performance over extended periods. Controller temperature rise, actuator load, battery level, sensor drift, and communication stability can all become decisive factors in the later stages of the run.

2. Why Autonomous Navigation is a Core Test

The competition encourages autonomous navigation and imposes stricter requirements on human intervention. This effectively distinguishes between the ability to “complete tasks via remote control” and “complete tasks autonomously by the robot.”

Autonomous navigation typically consists of perception, localization, mapping, path planning, and motion control modules. The robot first identifies track boundaries, obstacles, and directional markers using vision, depth information, or other sensors; then, combined with its own position estimation, it determines the current road segment; finally, it generates an executable motion trajectory and hands it over to the gait control system.

During actual operation, the navigation system must also face the following uncertainties:

• Light changes may cause fluctuations in visual recognition results;
• Ground texture and color changes may interfere with track boundary detection;
• Other robots or staff may temporarily occlude the area ahead;
• The robot’s own shaking can cause unstable image and sensor data;
• Localization errors may gradually accumulate over long running periods.

Therefore, reliable autonomous navigation should not rely on a single sensor or judgment condition. It should employ multi-source information fusion and anomaly detection. For example, when visual localization results suddenly jump, the system can perform consistency checks using inertial information and historical trajectories; when it cannot confirm if the road ahead is clear, it should prioritize slowing down or briefly stopping rather than continuing to execute uncertain actions at high speed.

3. System Engineering Constraints from Limited Robot Swapping Rules

The rules allow a maximum of two robot swaps throughout the race, with different time penalties applied. This design prevents teams from endlessly relying on equipment replacement to avoid failures, forcing them to balance robot reliability, operational strategy, and maintenance efficiency.

From a system engineering perspective, swapping strategies involve at least three types of decisions:

1. Preventative Swapping

If the robot exhibits risks such as continuous temperature rise, gait abnormalities, or insufficient energy, the team can proactively decide whether to swap devices. Preventative swapping can avoid severe failures but consumes limited swap opportunities and incurs a fixed time penalty.

2. Post-Failure Swapping

Swapping after the robot falls, loses control, or encounters critical component anomalies maximizes the operating time of a single device. However, this may result in longer downtime or even missing sectional cutoff points.

3. Task Segmentation Strategy

If a team possesses multiple robots in different conditions, they must also consider the adaptability of different devices to flat ground, slopes, and curves. The ideal approach is not simply running the highest-performing device for the entire race, but formulating a dynamic strategy based on the remaining distance, equipment health, and track difficulty.

The foundation of these decisions is establishing comprehensive operational status records, including battery level, temperature, joint load, fall count, sensor anomalies, and manual intervention events. Without traceable data, swapping decisions easily devolve into empirical guesses, making post-race analysis difficult.

4. Higher Energy Management Demands from Battery Swapping Rules

The competition stipulates that in-station battery swaps must be completed under referee supervision. While no additional time penalty is applied, the time consumed during the swap counts toward the total score. Out-of-station swaps are restricted to emergencies like sudden battery failures and incur progressively increasing time penalties.

This rule elevates energy management from “is there enough power?” to “how to reliably recharge within a limited timeframe.”

Robots need to establish an energy consumption model before the race, estimating at least the following variables:

• Energy consumption per unit distance across different terrains;
• Additional power consumption from uphill walking and frequent turning;
• Impact of low battery states on actuator performance and control stability;
• Impact of battery temperature changes on available capacity;
• Fixed time and safety inspection time required for the battery swap process.

Simply deciding whether to swap based on remaining battery is insufficient. Since actual energy consumption is affected by gait, speed, ground friction, and ambient temperature, the system should continuously calculate the remaining travelable distance and set a safety margin. When the predicted result falls below the safety threshold, it should initiate the swap process in advance to avoid being forced to stop mid-track due to power depletion.

5. Graded Handling Mechanisms for Falls and Collisions

Falls are not entirely avoidable for humanoid robots operating outdoors over long distances. The key is whether the system can reduce fall probability, recover quickly, protect the equipment, and avoid affecting other participants.

A complete fall handling process typically includes:

1. Fall Recognition: Determining if the robot has lost balance via attitude angle, inertia changes, joint positions, and foot contact status.
2. Motion Protection: Upon confirming a fall is inevitable, reducing joint impact to prevent actuators from hitting the ground in a high-stiffness state.
3. Environment Confirmation: Assessing whether there are personnel, other robots, or obstacles nearby.
4. Self-Righting or Manual Handling: Determining the permitted intervention method according to race rules and logging the event time.
5. Status Review: Re-examining sensors, joints, energy systems, and communication links to confirm the device is fit to continue running.

Collision handling should also adopt a graded mechanism. Minor contact can be resolved by decelerating and adjusting direction; continuous contact should trigger an immediate halt to prevent devices from crushing each other or causing greater risks. For organizers, track boundaries, buffer zones, staff positioning, and unified retrieval procedures are equally important components of the safety system.

6. Cutoff Mechanisms Test Stable Efficiency

The competition sets sectional cutoff times and a unified retrieval system, with a total cutoff time of 3 hours and 40 minutes. Multiple checkpoint locations also have independent cutoff points. Teams that fail to pass on time are marked as DNF (Did Not Finish).

This mechanism prevents equipment from lingering on the track and ensures that results better reflect continuous operational capability. Even if a robot possesses a high peak speed, frequent falls, repeated stops, or multiple manual interventions may prevent it from passing subsequent cutoff points.

Therefore, speed control should shift from “as fast as possible” to “maintaining stability within safe limits.” Engineering-wise, the total task time can be broken down into:

• Actual walking time;
• Turning and deceleration time;
• Fall recovery time;
• Battery swap time;
• Manual confirmation and safety check time;
• Other abnormal downtime.

Through this approach, R&D teams can identify the true sources of time loss. Many systems do not lack walking speed; they waste significant time in turning, localization loss, energy decisions, and fault recovery.

7. How Participating Systems Should Conduct Pre-Race Validation

For complex tracks, pre-race testing cannot be limited to repeating straight-line walks on flat ground. A test matrix covering real-world risks should be established.

Environment Dimension

Test flat ground, slopes, curves, narrow passages, and various ground connection areas separately. Focus on observing whether gait mutations occur during terrain transitions.

Task Dimension

Cover the complete workflow: starting, accelerating, decelerating, turning, temporary stopping, re-localization, battery swapping, and fault recovery. Do not validate only single actions.

Fault Dimension

Simulate short-term sensor anomalies, localization fluctuations, battery drops, communication delays, and local occlusions. Check if the system can enter safe mode rather than continuing to execute erroneous paths.

Operations Dimension

Verify referee supervision, the robot swapping process, data logging, and manual handling procedures. The technical system must ultimately run under competition rules; any unrecorded or unexplained intervention may affect scoring and post-race analysis.

8. Insights from Competition to Real-World Applications

The significance of the humanoid robot half-marathon extends beyond the competition itself. It provides a relatively clear comprehensive testing scenario to observe whether robots can continuously complete tasks in open environments.

For R&D teams, the most noteworthy aspects are not the single fastest times, but the following metrics:

• Success rate of long-term operation;
• Gait consistency across different terrains;
• Autonomous navigation stability;
• Recovery speed under abnormal states;
• Energy consumption per unit distance;
• Frequency of manual intervention;
• Data logging and fault tracking integrity.

In the future, as robots gradually enter scenarios such as inspection, logistics, public services, and complex operations, systems must evolve from “being able to execute actions” to “understanding the environment, running continuously, and recovering safely.” The emphasis on autonomous navigation, limited resources, and unified safety management in the competition rules represents the most common and challenging issues in actual deployment.

From this perspective, the half-marathon is not just a speed competition, but a comprehensive verification centered on reliability, adaptability, and engineering maturity. Those who can truly integrate perception, control, energy, software, and operations are more likely to build sustainably operating humanoid robot systems in real-world environments.