Platform

The complete autonomous training platform

From robot model to deploy-ready policy. One integrated environment for simulation, training, evaluation, and export.

Model Studio

Import, inspect, articulate

Drop any URDF, MJCF, or SDF robot model into the platform. The Model Studio automatically parses the kinematic chain, identifies joints and actuators, and generates an interactive 3D view.

Joint Visualization

Every revolute and prismatic joint rendered with range-of-motion indicators and real-time angle readouts.

Kinematic Analysis

Forward and inverse kinematics computed on import. Full Jacobian matrix available for each end-effector.

Collision Meshes

Toggle collision geometry overlays. Detect self-collisions before they become training artifacts.

Actuator Profiles

Torque curves, velocity limits, and gear ratios extracted and displayed per-joint. Ready for simulation.

Interactive joint control

Base0°
Shoulder-30°
Elbow45°
Wrist-20°

Training Engine

Tune, iterate, converge

Explore how hyperparameters shape the training curve. The reward autopilot tunes these automatically — here you can see why each parameter matters.

Hyperparameter explorer

05001k05M10MStepsReward
// PROTOCOL_ENFORCED: IACON-PRIMELearning rate3.0e-3
Batch size512
Entropy coeff0.010

The training engine runs thousands of parallel physics simulations on GPU. Each simulation instance generates experience data that feeds the policy optimizer in real time.

Learning Rate

Controls the step size of gradient updates. Too high causes instability; too low slows convergence. The autopilot uses a cosine annealing schedule.

Batch Size

Number of experiences per gradient step. Larger batches reduce variance but require more memory. Automatically scaled to hardware.

Entropy Coefficient

Encourages exploration by penalizing policy certainty. High entropy prevents premature convergence but slows final performance.

Architecture

Full-stack, purpose-built

Four layers from browser to bare metal. Data flows from user interaction through API orchestration into GPU training clusters and back.

System architecture

Browser
Editor
3D Viewer
Terminal
Training Dashboard
API Layer
Auth & Sessions
File System
Optimus Agent
WebSocket
Training Engine
Parallel Simulation
Reward Autopilot
Policy Optimizer
Evaluation
Hardware
Cloud GPU (A100/H100)
Network Storage
Container Runtime

Pipeline

From training to deployment

01

Training

GPU-parallel RL with autonomous reward iteration and multi-seed evaluation

02

Evaluation

Aggregate scoring across seeds with safety metrics and video rendering

03

Export

ONNX, PyTorch, TFLite — versioned artifacts with metadata and checkpoints

04

Deploy

Push to real hardware with domain randomization for sim-to-real transfer

Ready to build?

Get access to the full platform. Import your robot, train a policy, and deploy — all from your browser.