PLAYGROUND.md

SuperHQ Motion experiment: playground

Controllable motion. Runs in your browser.

This is a research experiment, not a released model. It shows how far a small learned locomotion model, combined with game-side runtime code and a few authored clips, gets toward natural game movement. Expect visible artifacts.

Controls

Input	Action
W A S D	Move relative to the camera
Shift / Ctrl + Shift	Jog / sprint
Space	Jump (standing, walking, jogging or sprinting)
C	Toggle crouch
F / G	Dash / lunge
H	Knockback test (pushed away from the camera)
Drag, scroll, hold right mouse	Look, zoom, shoulder view
P / R	Pause / reset

Open ? to switch the display character (Eric or the joint mannequin) and the display floor guard. Jumps and actions need the Natural style and standing height; Carrying and Zombie styles do not support them.

What is learned and what is code

Part	Source
Idle, walk, jog, sprint, crouch, carrying and zombie poses	Learned locomotion model (616,591 parameters)
Standing crouch and standing jump poses	Learned action model (699,053 parameters)
Running jump	Learned stride plus runtime code: waits for a support foot, plants it with IK, drives push-off and landing compression, time-warps the stride's flight. Not a learned running-jump policy.
Dash, lunge, knockback and get-up, ledge climb	Authored CC0 clips replayed and blended into and out of the learned locomotion
Turning, strafing	Code rotates the forward gait; no learned strafe, backpedal or pivot
Gravity, collisions, horizontal travel	Example host physics

The playground loads both model packages (1,315,644 parameters, about 5.26 MB of FP32 weights) and runs inference locally at 60 Hz.

Known limitations

• Running jumps take off 0.05–0.32 s after the press, because the jog stride has long airborne phases and takeoff waits for a foot on the ground.
• Feet slide during ordinary locomotion and in some transitions; turning is a code-driven rotation.
• Stopping from a run passes through a walk and takes about 1.5 s.
• Dash and lunge arms are posed for a shield and sword; the character holds nothing.
• The authored clips hold the head lowered by about 12° at idle and 20–23° when running. A constant 10.9° head-pitch conversion bias in the training data is removed before display.
• Ledge climbing needs a game-provided ledge and is not available in this flat playground.
• Development checks only; no blinded or user acceptance study.

Run locally

Serve the Space directory over localhost and open index.html. HTTPS or localhost is needed for weight checksum verification.

python3 -m http.server 8025 --bind 127.0.0.1

JavaScript

import {MotionGameController} from './src/game-controller.js';

const controller = await MotionGameController.load(
  new URL('./model/', location.href),
  new URL('./actions/', location.href),
  {clipsURL: new URL('./clips/action-clips.json', location.href)}
);

// At a fixed 60 Hz:
const pose = controller.step({moving: true, jog: true, heading: 0});
controller.requestJump();                       // once per input edge
controller.requestAction('dash');               // 'dash', 'lunge', 'knockback', 'climb'
controller.requestAction('knockback', {direction: [1, 0]});

rootPosition and localQuaternions describe the canonical pose. Apply worldPosition and heading once as the parent transform. For display, pass the pose through correctHeadPitch from src/pose.js.

A game can supply its own physics host with reset(), state(), launch() and step(); place(height) is needed for climbing. An optional resolveHorizontal callback passed to step() resolves travel against game geometry. rawComposedPose and blendCorrection expose runtime corrections.

Credits and licenses

Weights: CC BY 4.0. Runtime code: MIT. Locomotion data: 100STYLE (CC BY 4.0) and Quaternius Universal Animation Library 1 and 2 (CC0). Character: "Eric Rigged 001" by Renderpeople (CC BY 4.0). See ATTRIBUTION.md.