Robo-Bench: Gemini 4 Argon scores 0.83% on humanoid robot control
Vals AI's Robo-Bench has frontier models control a Unitree G1 humanoid in real time; Gemini 4 Argon scored 0.83%, and most models failed by thinking too long to send commands.
THE SMARTEST AI MODELS IN THE WORLD WERE GIVEN A HUMANOID ROBOT BODY NONE OF THEM SCORED ABOVE 2 OUT OF 100 ok wait this is from robo-bench by vals ai, updated october 13 they took frontier models, the ones that write code and pass exams and asked one question can you control a robot in real-time physics using only motor commands and camera images ngl the answer is humbling ▹ the setup a simulated unitree g1 humanoid: 29 joints, two dexterous hands a head camera, two motion sensors no built-in walking, no balance helper the model has to write its own controller, live three minutes per task, and the clock keeps running while it thinks 11 humanoid tasks, plus 13 tasks with a simpler robot arm ▹ the scores best overall: gpt-6.1 sol at 28.33% then gpt-6 astra 23.33%, claude opus 5.5 14.58% wonder full 10.42%, claude fable 5.1 5.83% gemini 4 argon 0.83%, grok 4.7 0.21% on the arm, the best model averaged 51/100 and solved 4 tasks on the humanoid: zero tasks solved no model averaged above 2/100 ▹ the funniest and saddest detail thinking too long is a failure mode the physics doesn't wait for you grok 4.7 sent no motor commands in 16 of 24 episodes gemini 4 argon in 15 they were still planning while the robot just stood there claude opus 5.5 got to 95 on a balance task then declared success before the beam held level for 2 seconds ▹ what this tells you about girl humanoids the ai that talks to you is not the ai that walks a chatbot can flirt, remember your birthday, write you a poem put the same brain in charge of 29 motors and it falls apart that's why real robots use separate, specialised control systems and why every "fully autonomous girlfriend" clip deserves a second look ▹ the surprise language was the easy part balance, timing and touch are the hard part and the gap is not small: 51/100 on the arm, under 2/100 on the body the brain wins the conversation the body still wins the room so in the replies what would you trust first an ai that talks perfectly or a robot that walks perfectly
This post quotes another post: x.com/rimtoln/status/2110842028678672597