We Replaced SSH Tunnels and Terminal Windows with a Browser Dashboard for NVIDIA Isaac Lab

Training humanoid robots on cloud GPUs without ever opening a terminal. This is Isaac Monitor.

If you've ever tried to train a robot policy on a remote GPU server, you know the pain. You're juggling five SSH sessions, tailing log files in one window, running nvidia-smi in a loop in another, praying your tmux session doesn't die, and trying to remember which run was the one with 512 environments and the modified reward function.

We got tired of it. So we built Isaac Monitor — a full-stack, real-time web dashboard that puts your entire NVIDIA Isaac Lab and Gazebo Harmonic workflow into a clean browser interface. No SSH. No terminal juggling. Just open your browser and train.

The Problem: Isaac Lab Deserves Better Than a Terminal

NVIDIA Isaac Lab is genuinely impressive. It runs massively parallelised robot learning environments on GPU — training a humanoid to walk in minutes, not days. But the tooling around it? It's a command-line affair. You configure runs in YAML files, monitor training by watching reward numbers scroll by, and debug failures by grep-ing through logs.

For a researcher running one job, that's manageable. For a team managing dozens of training runs across a cloud GPU cluster? It becomes chaos fast. Isaac Monitor is our answer to that chaos.

What Is Isaac Monitor?

Isaac Monitor is a full-stack web application that runs alongside your Isaac Lab and Gazebo Harmonic installation on your cloud GPU server. It exposes a clean, real-time dashboard accessible from any browser on any device.

Tech stack: Rust (Axum) backend on port 3001 with REST and WebSocket APIs, React frontend on port 5173, and a Python bridge (gazebo_bridge.py) for gz-transport integration.

Architecture: Isaac Lab / Gazebo → Python Bridge → Rust Axum :3001 → React UI :5173

It's built for headless cloud servers — the kind that live in an AWS data centre with an NVIDIA A10G GPU and no monitor attached. We run it live on an A10G instance running Isaac Lab 2.3.2 with driver 580.126.09. There are 9 Isaac Lab tabs and 5 Gazebo tabs.

The 9 Isaac Lab Tabs

1. System — Pre-flight & GPU Monitor

Before you waste 45 minutes on a training run that fails because your Python environment is broken, Isaac Monitor runs a full pre-flight check. One click verifies: Isaac Lab binary availability, NVIDIA A10G VRAM (23 GB available), GPU driver version, disk space, and Python interpreter. Everything shows green or red. No ambiguity.

Alongside the pre-flight check, you get live A10G GPU telemetry streaming in real time: utilisation %, temperature (°C), power draw (W), and VRAM usage (current / 23 GB), all displayed with a rolling 2-second bar chart.

2. Dashboard

The training dashboard gives you a single-glance overview of everything running on the server: active Isaac Lab processes, live reward curves, and server health metrics. If training is diverging, you'll see it here before you'd ever catch it in a log file.

3. History

A chronological event log of everything that has happened on the training server — commands executed, training events, errors, system messages. Full context for debugging, all in one scrollable timeline.

4. Run History — 15 Training Runs Logged

Isaac Monitor maintains a full database of every training run. Current runs include: Isaac-Velocity-Rough-H1-v0 (running, 1 env, 1,000 iters), Isaac-Velocity-Flat-Unitree-Go2-v0 (completed, 16 envs, 1,000 iters), Isaac-Humanoid-v0 (running, 512 envs, 1,000 iters), and 12 more. Filter by task name, tag runs, compare reward curves across runs side by side.

5. Launch — GUI Training Launcher

No more editing YAML files or copy-pasting long CLI commands. The Launch tab is a GUI training launcher for Isaac Lab: task selector (Anymal-C, Humanoid, Unitree-Go2), RL library (RSL-RL or RL-Games), number of environments, max iterations, random seed, and headless toggle. As you configure, a live command preview updates in real time showing the exact CLI command that will be run.

6. Robot Assets Browser

The Assets tab is a pre-loaded browser of robot model files: ant.xml (MJCF, 8-DOF), half_cheetah.xml (MJCF, 6-DOF), simple_pendulum.urdf (URDF, 2-DOF). The full pipeline is right there in the UI: URDF/MJCF → Isaac Sim Importer → .usd → Task Config → RSL-RL train.

7. Models — Checkpoint Browser

All your saved policy checkpoints live here. Browse by run, by task, by timestamp. Select a checkpoint and launch Sim Preview — an in-browser live simulation of your trained policy running in the simulator.

8. AI Chat — Powered by Model Context Protocol (MCP)

This is the one that genuinely surprised us with how useful it became. Isaac Monitor has a built-in AI assistant that connects directly to the training runtime via MCP tools. It reads live GPU stats, inspects training logs, compares run histories, launches jobs, and manages robot assets — all from natural language chat.

Example queries it handles in production: "Give me a full training snapshot — GPU, status, and latest rewards" → returns live GPU utilisation + reward mean for each run. "Why did my last training run fail?" → pulls the relevant log segment, identifies the cause. "Analyse my latest run. Are rewards plateauing?" → reads reward history, returns analysis. "List all saved models we have so far" → returns checkpoint inventory with metadata.

Four query categories: Live Status, Diagnose & Fix, Analyse & Compare, Control & Assets. It's not a chatbot layered on top of the UI — it has real tool access to the training runtime.

9. Sim Preview

A dedicated tab for in-browser live simulation preview. Load any saved policy checkpoint and watch it run. No VNC. No remote desktop. Just a browser tab.

The 5 Gazebo Harmonic Tabs

Isaac Monitor also provides full monitoring and control of Gazebo Harmonic simulations.

1. Overview

Shows simulator state (stopped/running), the loaded world file, active topic count, PID, simulator version, WebSocket bridge address (ws://localhost:8767), REST API endpoint, a real-time Real-Time Factor (RTF) chart, and the active topics panel. Data flow: gz sim → gz-transport → gazebo_bridge.py → Rust Axum :3001 → React UI :5173.

2. Topics

Once Gazebo is running, the Topics tab shows a real-time list of every active gz-transport topic — latency, message type, publish rate. Full observability into what the simulator is doing.

3. Spawn Models

Drop URDF or SDF robot models directly into the running Gazebo world from the browser. No terminal. Select a file, configure the spawn pose, hit spawn.

4. Worlds

A world file manager for Gazebo Harmonic. Switch between SDF worlds without stopping the simulator.

5. Logs

Streamed Gazebo simulation logs: Real-Time Factor, physics constraint errors, and topic activity. All searchable, all in real time.

The Numbers

Who Is This For?

Robotics research teams running Isaac Lab on cloud GPU infrastructure who need more than a terminal to manage training runs. Teams where multiple engineers need visibility into what's training, what completed, and what failed — without all SSH-ing into the same box.

ML engineers deploying robot policies who want to see their checkpoints, run ablation experiments, and compare reward curves without switching between tools.

Robotics startups who want a professional dashboard for their GPU training cluster without building one from scratch.

Deployment

Isaac Monitor is deployed on your own cloud infrastructure or on-premise GPU cluster. It runs on any server where Isaac Lab and/or Gazebo Harmonic are installed. The Rust backend is lightweight, and the React frontend is a static build served from the same machine. Custom deployment is available for teams.

Get Access

Isaac Monitor is currently available for teams under a custom deployment model. Contact us at support@istrux.tech with subject: Isaac Monitor Access Request or Isaac Monitor Deployment Enquiry.

Built by the team at Istrux. We're obsessed with making cloud robotics infrastructure as frictionless as possible.

← Back to Robosynx