CHALLENGE RESULTS

GPT-6 Astra Challenge

Thanks to everyone who built and submitted a robotics project with GPT-6 Astra. Congratulations to our winner and runner-up!

Winner
OpenArm Mona Lisa — LeRobot ACT
Runner-up
Fridge Speller
Submissions closedGet the build prompt
Prizes
$500 each · 2 winners
What to build
A robotics project built with GPT-6 Astra
Robot
Any robot — your choice
Original deadline
September 16, 2026 · 5:00 PM PST (UTC−8)
A small robot painting a portrait beneath a sunlit stone arch in Florence.

ANY ROBOT. YOUR IDEA.

The challenge

Builders chose any robot and built something interesting with GPT-6 Astra, from painting a portrait to playing a sport.

  1. Pick a robot. Build with GPT-6 Astra.Choose your robot and the skill or creative idea you want to explore.
  2. Show it workingCapture a short demo of your robot project and explain what Astra helped you create.
  3. Share the resultEntries included a project ZIP and demo MP4. Submissions are now closed.

START BUILDING

One prompt. Your next idea.

The original challenge prompt is available below for reference. Astra submissions are closed; you can still use the ideas for your next project.

Read the full build prompt
Help me build an entry for the GPT-6 Astra Challenge on HIM Arena: a robotics project built with GPT-6 Astra, using any robot. Two winners receive $500 each.

Work with GPT-6 Astra as the building assistant. If this session is not using GPT-6 Astra, tell me to select it before continuing; do not invent model access or claim another model is Astra. Pick any robot and build an interesting skill, simulation, or creative robotics project around it. Painting, playing a game, manipulating objects, and sports are all possible starting points. This is a robotics challenge; a general-purpose app or game without a robot is outside its scope. A sports theme and a runtime AI integration are optional.

Read the current repository instructions and inspect the project before changing files. Use the robot and idea I have supplied or that the repository already supports. Otherwise, choose a robot with an available, appropriately licensed model and one small, interesting robotics idea that fits this workspace. Tell me which robot you chose and what you are making. Ask only for details that block useful work. Preserve existing work and use the project's current tools where practical.

Build the smallest complete version with one clear moment worth showing. Run it, test the main interaction and a meaningful failure case, and fix problems you find. Use actual output and describe limitations honestly. Distinguish rendered concepts, scripted behavior, trained policies, and real hardware whenever relevant. Do not fabricate functionality, scores, training, or results.

Record a short result.mp4 showing the chosen robot and what the project makes it do. Capture the actual simulation, creative output, or existing hardware footage and label it accurately. Inspect the video and representative frames. If recording is unavailable, finish the build and explain exactly what remains to capture. Do not operate physical hardware without separate authorization and the project's safety procedures.

Prepare a README with the project name, chosen robot and model source, what it does, simulator or creative tools used, setup and run instructions, test results, limitations, licenses, and a short "Built with GPT-6 Astra" section describing what Astra actually helped create. Record the model name shown in this session when available; do not invent a model ID or require publishing private conversations.

Prepare submission.zip with reproducible source, dependency pins, the README, and only small required assets you have permission to share. Exclude .git, secrets, tokens, .env files, private conversations, datasets, caches, environments, logs, unrelated files, and the generated video. Keep result.mp4 at 50 MB or less and the ZIP plus MP4 at 100 MB or less.

The submission deadline is September 16, 2026 · 5:00 PM PST (UTC−8). Before uploading, show me the working project, exact run command, test evidence, video path, ZIP inventory and sizes, and a proposed public title and caption explaining the result and Astra's role. Ask for approval before public upload or paid compute. Do not push or deploy the project as part of entering this challenge.

After upload approval, use HIM Arena's Build page at https://arena.himrobotics.com/build?challenge=gpt-astra-challenge#upload, or the HIM CLI. Check authentication with him whoami --json. If login is needed, tell me to run him login and enter the developer token only in its hidden prompt; never request a token in chat.

Publish with: him upload ./submission.zip --video ./result.mp4 --kind [policy|simulator|both|creative-project] --robot [microduck|unitree-g1|agibot-x2|unitree-h1|unitree-go2|booster-k1|booster-t1|yam|so-101|franka-panda|other] --simulator [mujoco|isaac-lab|brax|custom|not-applicable] --challenge gpt-astra-challenge --name "YOUR PROJECT NAME" --caption "YOUR ROBOT, RESULT, AND GPT-6 ASTRA CONTRIBUTION" --run-command "YOUR VERIFIED COMMAND" --yes --json

Replace every placeholder with the truthful value before running the command. Any robot is welcome: use other for robots outside the listed models and name the robot in the README and caption. Use creative-project for creative robotics work that is not policy evidence, and not-applicable when no simulator is used. Return the published Arena URL and upload ID only after the CLI confirms them. If submissions have closed or publication fails, preserve the prepared files and report the error; do not publish under another challenge to bypass the deadline.

JUDGING

Make something worth watching.

HIM selected the most interesting submissions based on originality, technical execution, clarity of the demonstration, and reproducibility where relevant. Explain what GPT-6 Astra helped you build and show the result honestly.

PUBLIC SUBMISSIONS

All submissions

0:000:00
OpenArm Mona Lisa — LeRobot ACT
OpenArm Mona Lisa — LeRobot ACT

LeRobot ACT learned joint policy for OpenArm v2 right. 384 contact dabs with six palette wells in MuJoCo. Seven CAN tests, five configuration tests and a full one-dab runner cycle passed over virtual CAN. Video shows simulated arm motion. Physical arm and real paint NOT tested; hardware commissioning required. Scripted paint sequence, learned motor control.

e's profile photoeOther robot
0:000:00
OpenArm Mona Lisa — Physical Palette Runner
Work in progress
OpenArm Mona Lisa — Physical Palette Runner

Continued as OpenArm Mona Lisa — LeRobot ACT: https://arena.himrobotics.com/uploads/source_59ffd3291f3623fc498d . The new version includes a learned joint policy and an articulated simulator video. This older ROS 2 archive has scripted motor primitives and a motion-plan diagram; its physical-arm validation is still incomplete.

e's profile photoeOther robot
0:000:00
Fridge Speller
Fridge Speller

A Unitree G1 with Dex3-1 hands walks to a refrigerator, selects five alphabet magnets, and spells HELLO across the middle of the door! Trained in HIM Arena and built with GPT-6 Astra, the MuJoCo controller coordinates balance, finger contact, magnetic detachment, carry and release. The continuous nominal-scene run completes all five placements with five-second released retention checks. Fun!

Maddie D. Reese's profile photoMaddie D. ReeseUnitree G1

Results have been announced and submissions are closed. See Arena’s Challenge Rules for general terms.