Is That Humanoid Demo Autonomous or Teleoperated? How Race Rules and Videos Tell You

Every time a humanoid runs or dances on video, someone asks whether a person is driving it. This article starts from a race where autonomous and teleoperated robots shared one course under published rules, explains how the two modes differ, and lists what to check in demo videos and press material.

Teleoperation vs. autonomy: who decides what

“Teleoperated or autonomous” is rarely a clean either/or. In practice the human and the robot split the decisions, in different proportions.

At any of these levels, keeping balance while walking or running fast is the job of the robot’s locomotion controller, because nobody can move joints step by step at running speed. Our article on humanoid bipedal locomotion control covers how those controllers work.

What Beijing’s 2026 half-marathon rules showed

The second humanoid half-marathon in Beijing’s E-Town district, held on April 19, 2026, put autonomous and teleoperated robots on the same course. CGTN reported more than 100 teams, about 40% of them autonomous. Finishing times were weighted: 1.0 for autonomous navigation and 1.2 for remote control.

According to the Beijing municipal government, a pre-race test covered the full official distance of 21.0975 km across urban roads and park sections, with robots from both categories running side by side.

The weighting changed the outcome. CGTN reported that the first robot across the line was teleoperated and finished in 48:19, while the title went to an autonomous robot with a net time of 50:26. The arithmetic:

def to_seconds(m, s):
    return m * 60 + s

remote_raw = to_seconds(48, 19)          # teleoperated, actual time
autonomous_net = to_seconds(50, 26)      # autonomous winner (coefficient 1.0)

remote_net = remote_raw * 1.2            # teleoperation coefficient 1.2
print(divmod(round(remote_net), 60))     # (57, 59) -> about 57:59

# A teleoperated robot needed an actual time faster than this to win
break_even = autonomous_net / 1.2
print(divmod(round(break_even), 60))     # (42, 2) -> about 42:02

Converting the published times into average speeds over 21.0975 km puts the hardware progress in perspective. The 2025 winner’s time and the human record come from the same CGTN report; the speeds are our arithmetic (distance ÷ time).

In one year the winning average speed roughly tripled, and the autonomous winner held about 7 m/s for the full distance. The Beijing government’s pre-race report adds that Unitree’s H1, about 1.8 m tall and 47–62 kg, reached a peak of 10 m/s in testing. Peak speed and sustained speed are different numbers, though: a race rewards holding pace over 21 km with batteries, motors and cooling, not a short sprint.

The teleoperated robot was about two minutes faster on the clock but about 7 minutes 33 seconds slower after weighting; to win it would have needed an actual time under 42:02 (times and coefficients from CGTN, arithmetic above). The weighting reads as a deliberate push toward autonomous navigation.

Why teleoperation is still hard

A teleoperated robot that runs a half-marathon distance (21.0975 km per the Beijing government) without falling is itself evidence of a capable locomotion controller. The operator typically sends speed and heading; where to place each foot and how to lean is computed by the robot every moment. On identical hardware, the gap between teleoperation and autonomy sits mostly in seeing the route and deciding what to do.

So it is more accurate to read a teleoperated demo as proof that the hardware and locomotion controller work, and an autonomous demo as proof that the robot also makes the decisions.

How teleoperation data becomes autonomy

Teleoperation is not the opposite of autonomy; it is often the road to it. The OmniH2O paper uses human pose as a universal interface for real-time whole-body teleoperation through a VR headset, verbal instructions or an RGB camera, and reports learning autonomous policies from the teleoperated demonstrations it collects.

Robot arms follow the same pattern: a person drives a leader arm to record data, and a policy is trained on it. Our LeRobot SO-101 ACT training guide walks through that loop, and how many demonstrations you need covers dataset size.

Reading autonomy from a demo video, step by step

A video alone cannot prove autonomy, but this order gives you grounds for a judgment:

1. Check the labels first. Look for terms such as autonomous, teleoperated or real-time in the description or press release. If there are none, withhold judgment.
2. Look for people off to the side. Watch for someone walking alongside, a handheld controller, a VR headset or motion-capture gear.
3. Check speed and editing. Fast motion without a stated playback speed, or frequent cuts, hides real performance.
4. Find the attempt count. See whether the number of attempts and successes for the same task is published. One success is not a success rate.
5. Read the rules. For competitions, check how autonomous and teleoperated entries are separated and whether times are converted.

What to ask a manufacturer before a pilot or partnership

If a demo leads you toward deployment or joint development, these materials matter more than the video:

• The control mode used in the demo (whole-body teleoperation, high-level teleoperation or autonomy) and how often a person intervened
• Repeated-trial results for the same task: attempts, successes and failure types
• The sensors and compute that the autonomous features require

For turning spec-sheet numbers into task requirements, continue with how to read a humanoid robot spec sheet.