motornet.js — biomechanical motor control in the browser

Live demo

Watch a trained policy reach — right here

motornet.js · interactive demo

Click inside the canvas to set a reach target. A GRU policy trained in PyTorch drives the point-mass muscles to reach it — running here in plain JavaScript. Open the demo full-screen ↗

What it does

A faithful re-implementation of MotorNet's simulation engine

Effectors & muscles

Point masses and two-joint arms driven by Hill-type, Thelen, MuJoCo, and ReLU muscle models — the same biomechanics as the Python toolbox.

Gym-style environments

A RandomTargetReach environment with the familiar reset / step / obs / reward loop for closed-loop control.

Run trained policies

Export a PolicyGRU from PyTorch and run inference in the browser — no Python, no server, no WebAssembly.

Numerically validated

Trajectories are checked against the PyTorch reference across seven effector/muscle configurations plus the GRU policy.

Quick start

Simulate an arm in a handful of lines

import { RigidTendonArm26, RigidTendonHillMuscle } from 'motornetjs';

const effector = new RigidTendonArm26({ muscle: new RigidTendonHillMuscle() });
effector.reset({ jointState: [0.5, 1.0] });        // shoulder, elbow (radians)

const action = [0.3, 0.0, 0.6, 0.0, 0.2, 0.1];      // 6 muscle excitations in [0,1]
for (let t = 0; t < 50; t++) effector.step(action);

console.log(effector.states.fingertip);  // [x, y] endpoint position

See the README for environments, trained-policy inference, and the full API.