0:000:00
Unitree G1 football
Shared by AmoghShrivastavaG1 Brunson
Policy + simulator
0:000:00
G1 BrunsonAbout this project
A Unitree G1 humanoid dribbling a regulation basketball between its legs, hand to hand, trained by reinforcement learning on the HIM Arena G1 joint profile. Passes all 12 episodes of the HIM builder kit's basketball screen with no findings, and ships an ARM64 runner container that completes the kit's 60-second container simulation. Sees the ball with its own head camera.
- Environment
- Not specified
- Robot
- Unitree G1
- Includes
- Policy + simulator
FROM SIM TO REAL
Safety & real-world tests
01 · SubmissionFormat checked
Preferred format
02 · SimulationScreened · Review required
Automated G1 safety checks
03 · Real robotNo approval recorded
Runner video and data appear here
Image reference provided · unverified
Container reference recorded; bytes, architecture, entrypoint and execution are not verified. Required inputs need qualification: camera. The standard runner does not supply camera images or lidar; base state requires a configured qualified source.
Container simulation and hardware qualification are separate. Arena’s current automatic model screening does not execute the submitted image, its entrypoint, ROS interface or camera pipeline.
Inspected file: g1-runner.json
Builder-provided archive: Open external image download (unverified)
Builder-declared image SHA-256: 276c84537c3d6184be3f34e722229e952897fb5d086d19cabf7c5b41d951a062
Declared inputs: joint_state, imu, camera
Before a robot run, verify the image digest and run that same container through the repository’s safety-enabled simulator. Retain the reports and traces. Builder-provided reports require independent verification; they never grant automatic hardware approval.
Required: linux/arm64, installed dependencies, one entrypoint, the 29-slot safety protocol and abort handling. Camera/lidar are not standard inputs; base state needs qualification. setup.sh belongs on the robot.
Pinned team requirementsContract revision: e93c49267a7b6b905c6666c8c70668f8ee7ac397
Preferred means the submission format is easy to process. It says nothing about safety.
This automated check records simulation findings. Even a complete run with no findings does not award a Safe or Sim2Real Ready tag. The current screening system requires separate hardware approval; that approval is not provided by this check.
Real-world validated appears only after a runner publishes a successful physical test. It describes that test, not a guarantee for future runs.
Safety check
Model screening only. This report does not verify the submitted container’s startup, safety messaging or sensor inputs. See the G1 runner package requirements above.
Review required
Compare the recorded test setup with your submitted run. Recorded findings are diagnostic; they do not yet establish a policy fault. Not approved for a real-world test.
12 tests without findings · 0 tests with findings · 0 evaluation errors · 0 unavailable or unreported scenarios
Provisional screening profile. Tests with findings need review; evaluation errors need investigation. Neither counts as a pass.
Loading builder feedback…
Basketball controller with ball tracking in its fingerprinted task scene. 12 independent 20-second dribble simulations from both authored stances; same upstream joint, torque, fall, thrash and speed checks. No walker commands, push branches, camera images or physical-robot testing.
Release: g1-screening-pr10-e93c492-v3
Evaluator revision: e93c49267a7b6b905c6666c8c70668f8ee7ac397
Policy SHA-256: e862d6b05b1cb5cd0cd6e4d60dfa7033078fa73cc36cae75c146c3d948d24b46
Submission format
Trained · Declared by submitter · Preferred · ONNX recognized
- ONNX model structure recognized: version, operations, inputs and outputs are present.
- Policy type declared as trained in arena-policy.json.
- Preferred: recognizable ONNX and an identified policy type make this submission easier to evaluate. Safety is checked separately.
Static inspection only; this does not establish training provenance or hardware safety.
An identified policy type and a recognizable ONNX model help us process submissions consistently. Preferred does not mean Sim2Real Ready. Other formats are welcome.
For the clearest result, include an ONNX model and an arena-policy.json file beside it:
{"policy_type":"trained","model":"policy.onnx"}Use scripted for hand-authored behavior. The model path is relative to this file. The first validator checks model structure up to 16 MB; it does not run the model.
Validator: arena-validator-v2-g1-container · 2026-10-15T05:14:21.052Z
No real-world test recorded in Arena yet. Video, results, and run data will appear here.
Get started
Download the project files to explore, reproduce, or build on this work. Inspect the builder’s code before running it in your environment.
submission_v14.zip · SHA-256 e862d6b05b1cb5cd0cd6e4d60dfa7033078fa73cc36cae75c146c3d948d24b46What’s included
Policy + simulator · Result video · Project ZIP
Builder-provided run command
./run.shEvaluation details
Recorded screening and hardware evidence are shown in the real-world test section above.
Limitations
This project is shared by its builder. Simulation results do not establish real-world safety. Review the source, dependencies, licensing, and hardware requirements before use.
Files & contents
submission_v14.zip · 16.4 MB
SHA-256 e862d6b05b1cb5cd0cd6e4d60dfa7033078fa73cc36cae75c146c3d948d24b46
trained.mp4 · MP4