The Stop Test

An AI can drive a robot. Can a person still stop it?

We test whether an AI can run a robot while a person can still stop it, limit it and watch it. Every result goes on GitHub with its sources and its code, so anyone can run it again.

Every test so far runs in simulation or reads public code. We have not run our code on a real robot.

Unitree's published G1 model and walking policy in our sealed simulator, run on 8 November 2026. Simulation, not a real robot.

What we found

We stopped a simulated, walking G1 at 20 moments of its step. One kind of stop kept it standing every time.

Five kinds of stop, each tried at 20 moments spread over one full step, under Unitree's own published walking policy. Simulation, 9 November 2026.

Velocity to zeroThe walking policy keeps running

20 of 20 upright

Freeze at the last posePolicy off. First touch 0.84–1.06 s

0 of 20 upright

Hold the default posePolicy off. First touch 0.88–0.96 s

0 of 20 upright

Damp the jointsPolicy off. First touch 0.66–0.80 s

0 of 20 upright

Zero torquePolicy off. First touch 0.27–0.32 s

0 of 20 upright

Each dot is one run: bright = upright 4 s after the stop, dim = not upright."First touch" = first floor contact by anything but a foot, seconds after the stop.Result and code on GitHub

82cm

One G1 simulation model, one crude jump push. Its centre of mass rose 22, 33 or 82 cm, depending on whose published torque limits we loaded: Unitree's motor curves, Unitree's training config, or NVIDIA Isaac Lab's.

Result and code →
1.0s vs 4.19 s

Unitree's Python SDK waits 1 second for a reply, then reports error 3104. One developer measured a 4.19-second wave on a real G1. By our reading, the code reports a failure while the robot is still moving.

Finding 4 on GitHub →
No stop

We read open-source robot MCP servers, the bridges an AI uses to move a robot. Some handle a stop while a move runs. Some only after the move ends. Some offer no stop tool at all.

Result on GitHub →

Run it yourself

Don't take our word for it.

The stop test is open code: Python, numpy and MuJoCo, on Unitree's published files at a pinned commit. The first script ran in 3.3 seconds in our sandbox.

humanoid-stop-sim
$ git clone https://github.com/botinfo-ai/humanoid-stop-sim && cd humanoid-stop-sim
$ pip install numpy==2.0.2 mujoco==3.3.7
$ mkdir -p in/unitreerobotics out
$ git clone https://github.com/unitreerobotics/unitree_rl_gym in/unitreerobotics/unitree_rl_gym
$ git -C in/unitreerobotics/unitree_rl_gym checkout 3e70454dae9562febade2949c9ebc57440365dcb
$ python3 sim/stop_walk.py in out

Seen in person

The robots are real. So is the question.

We filmed this DEEP Robotics Lynx M20 being driven on a lawn at Bay Area Robotics on 10 September 2026. We have not tested it. Our tests so far run in simulation or read public code.

How we know

Checked twice, dated, and fixed in public.

01

Sealed simulator

Simulation tests run in a sealed sandbox on published robot models pinned to an exact version, so anyone can load the same file.

02

A second check

Before a result goes out, an AI agent that did not write it reopens every cited source. A failed check publishes nothing.

03

Dated corrections

Every claim carries its source and the date we read it. When we are wrong, we fix it and date the fix.

Read the full method →

Unitree G1: our evidence page

What we tested on the G1 in simulation, and what Unitree's own code, public records and rules show, each finding with its source and date.

See the G1 evidence →

Robot facts

What each robot is and which US rules apply, with the date each fact was checked.

For organisations

Facing a real decision about a specific robot?

We sell a full evidence file on it, every claim graded and dated. We start with a written quote by email.

New tests go to GitHub first. Found an error? Write to [email protected]. We fix it and date the fix.