TL;DR
Researchers successfully deployed a commercially available Unitree G1 humanoid robot to compete autonomously in 100m, 400m, and 1500m track events. The system overcomes the "real-world" hurdles of humanoid racing—such as violent camera shake and terminal motor overheating—by integrating stabilized vision-based lane tracking with a reinforcement learning (RL) locomotion policy and a proactive thermal management system.
What Problem Does This Paper Address?
While learning-based locomotion has enabled humanoids to run impressively in lab settings, taking them to a real competition like the World Humanoid Robot Games (WHRG) 2026 introduces a suite of "system-level" challenges.
First, the robot must be truly autonomous: it has to perceive lanes at an unfamiliar venue and steer itself without external help. Second, the physical act of running creates significant head-camera oscillations that break standard computer vision pipelines. Third, and perhaps most critically, sustained high-speed running generates immense heat. Without a strategy to pace itself based on hardware limits, a humanoid's motors will trigger emergency thermal shutdowns, leading to a fall and potential hardware damage.
Methods
The authors developed a dual-computer architecture (Figure 1) to split the workload between perception and control.
- Perception & Tracking: They used CLRerNet for lane detection. To handle the violent "head shake" during running, they implemented image stabilization using a Madgwick filter and IMU data. Detections are projected onto a ground plane and accumulated in a voxel grid to bridge moments when lanes are occluded or missed. Path following is handled by a Vector Pursuit algorithm.
- Locomotion Control: The running policy was trained in Isaac Lab using Proximal Policy Optimization (PPO). To ensure a natural, efficient gait, they used an Adversarial Motion Prior (AMP) derived from human motion capture data.
- Thermal Management: This is a standout feature of the paper. To prevent overheating, the authors:
- Penalized high torques during RL training.
- Added physical heat-dissipation plates to the actuators.
- Implemented a runtime thermal limiter that monitors MOSFET temperatures and automatically "caps" the robot's speed if it gets too hot.
Datasets
- Lane Detection: The model was fine-tuned on 3,000 real stadium images and 300 synthetic images. The synthetic images were rendered to match the specific indoor competition venue and made photorealistic using Google’s Gemini-3.1-flash-image model.
- Locomotion: Human motion capture data consisting of seven 50m sprints by two human runners was used for the AMP training stage.
Key Insights and Ideas
- Thermal Margin is the Bottleneck: In real-world racing, the limit isn't how fast the robot can run, but how long it can sustain that speed before melting its electronics. The thermal margin dictated the pace in every event.
- Synthetic Data for Venue Adaptation: Training on generic outdoor stadiums wasn't enough. The indoor venue had unique lighting and reflections; adding a small amount of venue-specific synthetic data was vital for robust lane detection.
- Stabilization is Mandatory: The head camera rolled by over 3° RMS during sprints. Without software-based stabilization and roll-leveling, the lane detector's performance degraded significantly.
- Bridging the "Flight Phase": Humanoid running involves flight phases where no feet touch the ground. The authors developed leg odometry that uses IMU integration to "bridge" these phases, maintaining an accurate position estimate for path tracking.
Key Results
The system was deployed on a Unitree G1 at the WHRG 2026 in Beijing with the following results:
- 100m Sprint: Finished 13th out of 74 (24.3s). The thermal limiter was active here, proving that even a short sprint pushes these robots to their limits.
- 400m Race: Placed 10th out of 44 in the first round (1:55.93). Unfortunately, the robot fell in the semi-final due to a lane-detection error.
- 1500m Endurance: Finished 7th out of 35 (7:43.89).
The quantitative breakdown in Table I shows that the robot maintained a mean leg-odometry speed of 4.26 m/s in the 100m, while pacing itself at ~3 m/s for the longer 1500m.
Limitations and Future Work
The most significant failure was the fall in the 400m semi-final. The authors suspect a hardware shift (a bent camera mount from a previous fall) changed the camera's pitch, causing the lane detector to fail. This highlights the need for perception systems that can self-calibrate or recover from unexpected physical changes.
Additionally, while the thermal management worked, the authors noted that modeling actuator thermal behavior in simulation is notoriously difficult. Future work could involve more sophisticated, learned thermal models to allow the RL policy to "feel" and react to heat more naturally during training.
Summary updated