HAGOP Robot Lab — Design / Train / Build

Import my robot · Rhino first

Open native Rhino .3dm, STL, OBJ or GLB locally on your phone. Inspect moving links, define joints/actuators and export an assembly specification for future physics validation.

OPEN CAD IMPORTER →

Printable course geometry

This produces a real, closed triangular-mesh STL for a printable test-course ramp or block. It does not yet generate a fully articulated robot from a text prompt; that requires a cloud CAD agent and mechanical validation.

Parametric geometry

Part type Units Length (X) Width (Y) Height (Z) Robot template

Shared sim/print dimensions

Use the exported dimensions when constructing matching obstacle objects in a MuJoCo or Isaac simulation. Simulation units must be converted from millimeters to meters.

Check printer bed fit, nozzle clearances, stability, and slicer geometry before producing parts. This reference shape is not a safety-rated load-bearing component.

Editable robot geometry

XGO-Duck printable structure ↗PCB, STEP, BOM and assembly ↗

Custom Rhino/Fusion robot geometry is handled in Import My Robot. XGO-Duck commands below train the XGO-Duck reference robot only, not arbitrary imported CAD. Imported assembly definitions must become a verified physics model and custom RL task first.

GPU training job

Task Parallel environments

Training launch command

OPEN RUNPOD ↗

These are setup commands, not a running cloud job. Read the full RunPod phone setup or inspect the guarded training helper. A RunPod GPU Pod must be started in your account, with Python 3.12, uv, compatible NVIDIA/CUDA drivers and a persistent volume. Start at 128–512 environments, then scale based on VRAM and measured throughput.

Training stack

XGO-Duck + mjlab is the recommended first real training route because the open-source project already has its robot geometry, Feetech servo actuator model, PPO tasks and ONNX exporter.

Training repository / documented commands ↗NVIDIA video workflow ↗Isaac Sim browser livestream ↗

Training → physical policy

Train and evaluate in the correct XGO-Duck simulation.
Export a compatible ONNX policy from a selected checkpoint.
Calibrate the assembled robot before any torque-enabled test.
Load the ONNX policy using its matching onboard runtime.
Compare actual IMU/joint data to simulation and retrain.

The original NVIDIA Microduck Isaac Lab policy is not automatically compatible with XGO-Duck. Hardware geometry, servo model, observation ordering and action interface must match.

Live training / simulation links

When you have a GPU Pod running, enter the authenticated endpoints here. Addresses stay in this phone's browser storage; no secrets are sent to hagop.io.

TensorBoard URL Simulation / WebRTC viewer URL

NVIDIA notes that bare Isaac Sim stream ports have no built-in authentication/encryption. Never expose them directly to the public Internet; use TLS, authentication and an appropriate private network or restricted reverse proxy.

What you'll see

01 Training curves — PPO reward, losses, steps, stability

02 Selected rollouts — successful walking, collisions, falls

03 Policy checkpoints — compare versions before hardware tests

04 Real-robot telemetry — future protected bridge with live video

The web panel does not invent a simulation feed. It opens streams only after you connect your own running service.

First physical build · XGO-Duck

This reference build uses an Arduino UNO Q, 15 Feetech HLS1910 serial servos, custom expansion board, QMI8658A IMU, printable structures and required power hardware.

The 2026-10-02 hardware repository warns that older board revisions reverse GND and SIGNAL on U3/U4/U5. Verify board revision, polarity, servo IDs and current protection before applying power. Test with the robot supported and torque off.

Official project resources

Bill of materials and power requirements ↗ 25-page illustrated assembly guide ↗ Printable STL components ↗ Arduino UNO Q firmware + browser control ↗ Train + export ONNX policies ↗

Control model

The robot's existing runtime controls servos through its MCU and exposes a local browser interface for calibration, inference, walk, recovery and torque-off. Future phone-over-Internet access needs an authenticated bridge and physical emergency-stop path.

What's not provisioned

No GPU has been started, no firmware has been flashed and no physical parts have been ordered or actuated.

Railway control plane

Existing authenticated Railway deployment provides robot templates and connection capabilities. Read-only endpoints work independently of RunPod.

CHECKING API…GPU NOT CONFIGURED

Checking endpoint…

RAILWAY PROJECT ↗RUNPOD CONSOLE ↗

Paid GPU creation stays disabled. Connect your RunPod and owner tokens using Railway Variables; this page can then show managed Pod status, request stops and display real training telemetry. Use provider auto-stop for the first trial.

Available robot profiles

Loading…

Backend integration

API health ↗Backend capabilities ↗Robot profiles ↗

Cloud execution is separate from on-device STL generation. A connected RunPod training worker will later run the upstream PPO task, store checkpoints and publish telemetry to this panel.

$5 pilot budget guard

For the first RunPod trial: at most 60 minutes, with a GPU priced at $1.50/hour or less. Allow $1 of headroom for storage and other charges. Verify the actual rate in RunPod before starting.

Quoted GPU rate ($/hour)Training window (min)

SAFE POD SETUP ↗

Waiting for rate verification…

This is a planning guard, not a provider-enforced spending limit. Use RunPod's --stop-after 1h feature and confirm the Pod stopped. A stopped Pod's volume can continue billing until it is terminated after you save its results.

Private RunPod + live PPO telemetry

Read managed Pod status, stop an XGO-Duck training Pod and watch metrics forwarded from its actual TensorBoard logs. Requires server-side RunPod and owner keys configured in Railway.

Owner access token (never saved to this website)

Managed Pod

Refresh every 10s

Private monitoring is off until you provide your owner token.

Actual reinforcement learning progress

No live training telemetry received yet.

No credentials are stored in localStorage, added to URLs or sent to GitHub. Only the owner token is transmitted in an authenticated header to your Railway API. The separate GPU reporter will send scalar values; no fabricated training curve is displayed.