G1 Football

Take five penalty shots against a simulated Unitree G1 goalkeeper driven by Metatron. MuJoCo physics, the keeper-mounted camera and ball perception run in your browser. This is a simulation game, not a real-robot controller.

Note

The compiled bundle is a public release (release_status: public-release). Its evidence is a bounded gameplay audit of 24 paired kicks and Chromium touch emulation; it is not a general save-rate estimate or a physical iPhone/Safari qualification. This is a simulation game, not a real-robot controller.

Playing

The demo is deployed at demos/g1-football/ beside the documentation. The page loads the pitch, vision runtime and Metatron in sequence and warms up the keeper in the background. The opening panel shows progress. Click Start game to play immediately when ready, or to queue the start while loading finishes. Readiness alone does not enable aiming or shooting. The keeper inset is labeled READY · PRESS START while waiting. If loading fails, Reload to retry restarts the page rather than partially restarting the runtime.

Use W / S to aim higher or lower and A / D to aim left or right. Hold Space to charge power, then release to shoot. On a touch screen, drag the aim pad and hold/release the charge button. Each round has five shots; use Play again after the round finishes. The keeper-camera panel shows the camera feed used by the ball detector, not a prerecorded animation.

Runtime and download

Opening the page downloads approximately 403 MB of unchanged FP32 policy weights from Metatron on Hugging Face. The immutable revision and SHA256 checks are recorded in policy/manifest.json. No account or token is needed. Small external-data graphs ship with the compiled application; the large weights do not.

Verified weights can be reused from browser storage when quota permits. Cached weights are checked again before use. Clearing site data, storage eviction or insufficient quota can require another download. The mobile loading path uses sliced cache verification and Blob URLs to reduce temporary memory copies, but the model and its runtime still require substantial memory.

Automatic provider selection uses a suitable WebGPU adapter when available and falls back to CPU/WASM. Software GPU adapters are not selected automatically. Use ?provider=wasm to select CPU inference explicitly. On a static host, a service worker scoped to this demo adds cross-origin-isolation headers and may reload the page once. Use HTTPS or localhost, not file://.

Known limits

Touch controls and mobile memory safeguards are implemented, but an actual iPhone/Safari device has not been qualified. Browser preview CI checks hosting, local perception, isolation and the background-preload request. It explicitly intercepts policy initialization to avoid downloading weights; it does not establish model readiness, mobile compatibility, gameplay quality or hardware WebGPU performance.

The keeper sees only its own camera. Wide low or mid shots (a camera-derived crossing at least 1.2 m to the side) trigger a goalkeeper dive: the keeper shuffles and drops toward the ball line as early as its body allows, holds the low pose until the ball arrives, then hands over to the native recovery. The dive follows a reference clip derived from one BONES-SEED motion, credited in the bundle’s NOTICE.txt. Nearer shots keep the native timed save, which paired camera tests showed is stronger there; high corners (ball above 0.9 m at the keeper) are not dived at because the keeper has no jumping reference. Over 24 paired kicks with identical camera noise, the dive saved 21 where the native save made 16: all six 1.5 m medium-pace kicks became saves, two far-post slow kicks were lost, near kicks were identical. Getting up after a dive is slower than after the native save. If the detector finds no plausible ball in its existing search window, it retries with a smaller image-closing kernel. This avoids merging a small ball with nearby white pitch markings. The retry uses the same camera pixels, shape checks and uncertainty thresholds; it does not receive the physical ball position or bypass trajectory confidence. The keeper can still miss high or very wide shots and lose balance. No public hardware-WebGPU speed claim is made. Performance depends on the browser and device; inference can run slower than real time.

Licensing

Applicable ProtoMotions-authored software retains Apache-2.0. This does not relicense supplied game or perception components or grant rights beyond their supplied notices. NVIDIA-provided exported models and weights are under the NVIDIA Open Model License; release packages must include the model card, full license and attribution notice. Runtimes and robot assets retain their own terms. See the bundle’s NOTICE.txt and licenses/ directory. Compiled browser assets remain downloadable and inspectable; compilation is not encryption or DRM.

See demos/g1-football/README.md for local install instructions, the evidence behind the release status and the dive reference’s provenance.