SuperHQ Motion (experiment)

Controllable motion. Runs in your browser.

This is a research experiment, not a production model. It is a small learned locomotion model for humanoid game characters, combined with game-side runtime code and a few authored clips. It is useful for studying a compact, browser-run motion controller; it is not yet natural enough to ship as a game's character animation.

Try it: SuperHQ Motion playground.

What is in this repository

Path	Contents
model/	Locomotion network: 616,591 parameters, 34 inputs, 143 outputs, FP32 (2.47 MB)
actions/	Action network: our own frozen 616,591-parameter base plus two 41,231-parameter residual adapters (stance, jump); 699,053 parameters, 48 inputs
clips/action-clips.json	Eight authored CC0 clips on the canonical rig (dash, lunge, knockback, get-up, ledge climb, jump references)
src/	Dependency-free JavaScript runtime: inference, game controller, running-jump composition, IK, clip playback, Three.js adapter
examples/, index.html	The playground, including the display character
evaluation/	Saved development evidence (see below)
provenance.json	Source checkpoints, training jobs, data licenses and SHA-256 of every file

Output: 23 six-dimensional joint rotations, a 3D pelvis offset and two foot contact logits for a 28-landmark, Y-up, metre-scale rig at a fixed 60 Hz. Both packages together: 1,315,644 parameters, about 5.26 MB.

What is learned and what is code

Behavior	How it is produced
Idle, walk, jog, sprint, crouch, carrying and zombie locomotion	Learned (locomotion network), conditioned on gait profile, trajectory preview and phase
Standing crouch and standing jump poses	Learned (action network) from synthetic key-pose targets
Running jump	Learned stride plus runtime code: support-foot takeoff timing, IK foot planting, push-off and landing pelvis curves, time-warped flight. Not a learned running-jump policy.
Dash, lunge, knockback and get-up, ledge climb	Authored clips replayed with offset blending into and out of locomotion
Turning and strafing	Code rotates the forward gait; no learned strafe, backpedal or pivot
Gravity, collisions, horizontal travel	The host game (an example host is included)

Usage

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

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

// Call once per 1/60 s.
const pose = motion.step({moving: true, jog: true, heading: 0});
motion.requestJump();                 // on an input edge
motion.requestAction('dash');         // 'dash' | 'lunge' | 'knockback' | 'climb'
draw(correctHeadPitch(pose));         // canonical pose; apply worldPosition and heading once

Serve the repository over localhost (python3 -m http.server) and open index.html to run the playground. Weight checksums are verified on load. See PLAYGROUND.md for controls, the physics-host contract and details.

Training

• Locomotion: our earlier 100STYLE-trained locomotion network, fine-tuned on retargeted Quaternius Universal Animation Library game loops in three short studies (game repair v3, floor refine v4, idle refine v5) on one NVIDIA L4; recorded compute about USD 0.49. This release is seed 431 of idle refine v5, fixed final update (1,000 updates), with no checkpoint selection on fresh cases.
• Actions: two residual adapters on our frozen base, trained on synthetic stance and jump targets built from authored key poses (action repair v11 stance, v9 jump; composed without further training).
• Inputs are the profile clock and trajectory preview; the networks do not consume the history of applied poses.

Evaluation

All results are development evidence on cases inspected during development; there is no independent test set or user study.

Software checks (passed): weight checksums; JavaScript and PyTorch agreement (model/browser-parity.json); 180 runtime tests; the playground loads and runs in desktop Chromium, locally and as the hosted Space (1.7 ms per step).

Learned behavior, measured:

• The locomotion network reproduces its source clips closely: head pitch within 0.3 degrees of the source at idle, walk, jog and sprint; development gates for idle and cycle pose error passed (evaluation/locomotion-gates.json).
• The action network's standing jump is learned from a few stationary key poses blended over fixed timings; its targets do not contain a real jump.

Running jump (runtime composition, 64 phase cases, evaluation/running-jump/):

Measure	Previous runtime	This release
Pelvis vertical-velocity jump at takeoff	2.8 to 4.1 m/s in one frame	0
Support toe height at takeoff (jog, median)	179 mm	3 mm
Landing foot ahead of pelvis	0.08 m	0.42 to 0.56 m
Blend floor lift required	up to 63 mm	0
Request to takeoff (jog / sprint, median)	0 ms	233 / 150 ms

These held on untuned phases, a different request frame and two other jump impulses.

Known failures and limitations

• Head pitch: source conversion pitched the canonical head 10.9 degrees forward; every checkpoint carries this bias. correctHeadPitch removes it for display. The source clips still hold the head about 12 degrees down at idle and 20 to 23 degrees down when running.
• Running-jump latency: takeoff waits up to 0.32 s for a support foot, because the jog stride spends about 70 percent of each cycle airborne.
• Foot sliding: ordinary locomotion, including gait transitions, slides up to about 3.3 m/s near the floor by our screen; sprint toe-off after landing up to about 6 m/s.
• Stops: stopping from a run routes through walking and takes about 1.5 s.
• No directional data: turning is code-driven; no strafe, backpedal or pivot.
• Actions: dash and lunge arms are posed for a shield and sword; climbing needs a game-supplied ledge and has no ledge alignment.
• Scope: flat ground, upright humanoid, the rig in model/profile-controller.json. Crouched and carrying jumps are not supported. Walking and standing jumps use the older key-pose jump prior and look stiff.
• The integration evaluator's "zero weight equals locomotion" and "canonical root step under 50 mm" checks encode the previous runtime design and fail by construction for the stride jump (evaluation/integration/).

License and attribution

Weights: CC BY 4.0. Runtime code: MIT (LICENSE-CODE). Training data: The 100STYLE Dataset by Ian Mason (CC BY 4.0); Quaternius Universal Animation Library 1 and 2 by Quaternius and Gonzalo Furnier (CC0). Authored clips: the same Quaternius libraries (CC0). Display character: "Eric Rigged 001" by Renderpeople (CC BY 4.0). See ATTRIBUTION.md.