← Paper Digest

Autonomous Track Racing with a Commercially Available Humanoid: System Design and Real-World Deployment

Koki Inami, Takashi Sato, Hiroyuki Oka, Shota Takizawa, Yuya Takahashi, Amon Suzuki, Hayato Ono, Kohtaro Ihara, Yuki Noka, Shohei Matsugu

November 2026

TL;DR A commercially available Unitree G1 ran the 100 m, 400 m and 1500 m fully autonomously at the World Humanoid Robot Games 2026, using head-camera lane detection, leg odometry and an RL running policy. It placed 13th of 74 in the 100 m and 7th of 35 in the 1500 m and fell in the 400 m semifinal; motor heating set the pace in every race.

What they do

They entered a commercially available Unitree G1 in three track events at the World Humanoid Robot Games 2026 in Beijing. The robot had to follow its lane on an unfamiliar indoor track with no input after the start command and no restarts. They added a Jetson AGX Orin, a RealSense D555 head camera and heat-dissipation plates (+1.5 kg). Because a fall could damage the robot, they ran no on-robot ablations. The paper reports the system, the race results and the failures.

How it works

  • Lane perception. CLRerNet is fine-tuned on 3000 frames from 30 outdoor stadiums plus 300 synthetic frames of a MuJoCo replica of the venue. IMU-based stabilization removes head roll (2.0-3.3° RMS). Detections accumulate on the ground plane, which fills in missed or occluded lanes.
  • Odometry and tracking. The G1’s built-in velocity estimate reverses sign during each step. Their leg odometry instead bridges flight phases with the IMU and measured 377 m on a 400 m practice lap. Vector pursuit with a 6 m lookahead sends speed and yaw-rate commands.
  • Running policy. It is trained with PPO in Isaac Lab, an adversarial motion prior from human sprint captures, and a squared hip-pitch torque penalty as a proxy for heating. Separate sprint and endurance policies were trained.
  • Thermal pacing. Above 120 °C the speed cap ramps toward 2 m/s, and at 127 °C the robot stops, just short of the 130 °C firmware cutoff. Each event gets its own speed cap.

Results

  • 100 m: 13th of 74 in 24.3 s. The temperature crossed 120 °C at 22.1 s and peaked at 123 °C.
  • 400 m: 10th of 44 in the first round (1:55.93). In the semifinal, lane detection looked abnormal at startup, and the robot veered right and fell about 5 s in.
  • 1500 m: 7th of 35 in 7:43.89 (seven entrants finished). The temperature peaked at 119 °C.
  • At least four lane lines were detected in every frame of the logged races. Before the thermal fixes, over-temperature protection stopped the motors after about 50 m.

Limitations

Without ablations, no component’s contribution is measured. The semifinal fall is unexplained because that run was not logged. At the venue, odometry distance came out about 0.90 of the true lap length. Heat limited speed in every race.

Figures

From Figure 2: Speed cap (top) and motor-driver temperature (bottom) from the logs of the three races, timed from the...
From Figure 2: Speed cap (top) and motor-driver temperature (bottom) from the logs of the three races, timed from the start command. In the 100 m the temperature crosses the 120 °C limiter trigger and the cap drops. In the 400 m the cap is raised for the last straight. Insets enlarge the first 30 s of the 100 m.
Abstract

Learning-based methods now let humanoids run robustly, and they have begun to compete in athletic events. Yet few reports describe a complete autonomous humanoid running system in an actual race, where the robot must follow a lane at a new venue, pace itself within hardware limits, and finish without intervention. We report such a system deployed on the Unitree G1, a commercially available general-purpose humanoid, in the 100 m, 400 m, and 1500 m events of the World Humanoid Robot Games 2026. The system integrates vision-based lane perception, path tracking, and a running policy trained by reinforcement learning with a human motion prior. To address the challenges of humanoid running, we stabilize head-camera images against body motion and use IMU-assisted leg odometry to bridge flight phases. To manage motor heating, we penalize hip-pitch torque during training, add heat-dissipation plates, and apply temperature-based speed limiting with event-specific pacing. The robot placed 13th of 74 in the 100 m, 10th of 44 in the 400 m first round before falling in the semifinal, and seventh of 35 in the 1500 m. In this deployment, thermal margin set the pace in every event, and the semifinal fall followed a lane-detection error.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.