R O C K E T A C A D E M Y
From planetary terrain to traceable robot experience.
Explore worlds. Assemble a team. Record what happens next.
Explore the worlds · Run locally · Read the vision
▶ Watch the concept film with sound
10 seconds · 720p · Original MP4
The film presents the platform vision. Implemented capabilities and research milestones are listed below.
FIRST CONTACT is a planetary robotics workspace built around a simple idea: every mission should leave behind an inspectable record of where the robots were, what they observed, and how the episode was produced.
Move from a planetary globe to a local terrain patch. Compare real robot platforms, prepare a scenario, run two rovers together, and replay the recorded outcome. MongoDB Atlas connects the scenario, source assets and saved recording; MCAP carries the episode evidence into other tools.
| Worlds | Robot catalog | Shared simulation | Episode evidence |
|---|---|---|---|
| Earth · Moon · Mars | 30 platform profiles | 2 independent robot instances | RGB · depth · states · sampled actions |
| Destination | Step onto the surface | Terrain foundation |
|---|---|---|
| MARS / DINGO GAP | Explore Curiosity's rocky passage in Gale crater. A selectable marker on the Mars globe opens the local reconstruction. | NASA/JPL colorized ground reconstruction, streamed as 3D tiles |
| MOON / APOLLO 17 | Traverse a 512 m patch in the Taurus–Littrow landing-site vicinity. | LROC NAC stereo terrain and matching orthophoto; 2 m elevation posts |
| EARTH / AMBOY CRATER | Build a terrestrial analog mission in the Mojave Desert. | USGS 3DEP 1 m elevation grid and USDA NAIP imagery; 512 m crop |
Pan, orbit and zoom through Three.js views. Choose a site from its globe marker or the world menu, then capture a complete 16 × 16 m terrain patch for Robot lab. Source identity, hashes and coordinate conventions travel with the episode.
Explore terrain provenance and coordinate frames →
01 — Choose your world. Open the globe, enter a measured site, and select the patch that will define the simulation surface.
02 — Assemble your team. Browse humanoids, quadrupeds, mobile manipulators, companions, drones and field robots. Compare up to four platforms with dated specifications and source links. The current physics pilot runs Clearpath Husky and Jackal; either model can occupy either robot slot. Unitree Go2 supports 3D inspection.
03 — Shape the scenario. Prepare mission intent in Scenario Studio. Save a prompt brief, optionally import world-model reference media, and freeze a version tied to a terrain site and robot configuration. Gemini also supports structured planning in Mission tools.
04 — Run, replay, retain. Generate a shared-world MuJoCo episode, inspect the recorded trajectories and camera previews, and download MCAP. Atlas stores the scenario lineage, metadata and recording hashes; GridFS retains the recording bytes.
flowchart LR
W["Choose terrain"] --> S["Freeze scenario"]
S --> R["Run robot team"]
R --> E["Record episode"]
E --> A["Atlas + MCAP"]
A --> V["Replay + inspect"]
Use Node.js 24, npm, and a browser with WebGL hardware acceleration.
git clone https://github.com/k2jac9/rocket-academy-first-contact.git
cd rocket-academy-first-contact
npm ci
npm run devOpen localhost:5173. The app opens at Dingo Gap; the local API runs on port 8787.
World exploration, the robot catalog and local kinematic mission tools work without service credentials. Globe images and the Moon/Earth site packages are bundled; Dingo Gap's streamed tiles need network access.
For saved physics episodes, configure MongoDB Atlas in .env.local and install the Python simulation environment. Add GEMINI_API_KEY for optional AI planning. Keep credentials server-side.
| Setup | Guide |
|---|---|
| Atlas, Gemini, optional terrain access, controls and API | Development guide |
| MuJoCo environment, robot assets and episode capture | Simulation guide |
| Connection diagnostics and verified recording reads | Atlas connection checks |
| Area | Current capability |
|---|---|
| Worlds | Earth, Moon and Mars globes; three local terrain sites; source-aware patch capture |
| Robot catalog | 30 profiles across six classifications; search, filters, 3D inspection where available, specification comparison and CSV export |
| Physics pilot | Two Husky/Jackal instances in one MuJoCo world; body-specific gravity; rigid terrain contact; simple virtual RGB/depth sensors |
| Scenario Studio | Exploration briefs, imported reference images/videos, provenance and immutable runnable pilot settings |
| Atlas | Saved runs, physical episodes and scenarios; hashed GridFS artifacts; exact filters and native semantic episode search (Preview); active connection probes |
| Replay and export | Recorded poses, camera previews, local recovery artifacts and compressed physical-episode MCAP |
The physical pilot is experimental. Sensor calibration, planetary contact models and downstream policy training are still research milestones. Catalog membership does not imply a runnable robot model; household robots and standalone arms are excluded from the catalog.
Google Genie: the app supports versioned exploration briefs and media imports. Live world generation and automated robot control through Genie remain unverified. Cosmos, Isaac Sim/Isaac Lab workers, governed training datasets and policy evaluation are proposed extensions in the platform design.
In Robot lab, select Mission audio for an editable briefing or saved-episode summary. Load your ElevenLabs voices, generate narration, then play or download the MP3. Audio is a presentation track, independent of the recorded robot timeline.
Set the server-only ELEVENLABS_API_KEY in .env.local, or run powershell -NoProfile -ExecutionPolicy Bypass -File scripts/configure-elevenlabs.ps1 on Windows for a masked prompt. Enable Text to Speech and Voices read permissions, restart the API and refresh the panel. Generation uses your ElevenLabs account credits; audio never plays automatically. See narration setup and limits.
| Layer | Implementation |
|---|---|
| Interactive workspace | React · TypeScript · Three.js · Vite |
| Local API and validation | Node.js 24 · shared TypeScript contracts |
| Physical execution | Python · MuJoCo |
| Mission planning | Deterministic baseline · optional Gemini |
| Mission narration | Optional ElevenLabs text-to-speech · editable scripts · MP3 playback/export |
| Episode memory | MongoDB Atlas · GridFS · SHA-256 lineage |
| Portable recordings | MCAP · explicit schemas, frames and timestamps |
npm test
npm run build
npm run check:atlasThe Atlas diagnostic checks the configured connection, indexes and representative recording hashes. Add -- --write for a temporary metadata/GridFS round trip with cleanup. See the verification report and visual QA for dated evidence.
| Explore | Read |
|---|---|
| Product architecture and the learning loop | Platform design |
| Terrain sources, registration and reproducibility | Surface sites |
| Genie boundaries and exploration provenance | Google Genie integration |
| Physics scope and remaining milestones | Synthetic episode lab |
| Shared interface language | Design system |
| Original film and preview details | Media notes |
| Mission audio and voice configuration | ElevenLabs narration |
| Atlas roadmap and delivery status | Sequential implementation plan |
| Native Atlas semantic-search prototype | Setup, operation and limits |
ROCKET ACADEMY · FIRST CONTACT
A place to explore. An episode to learn from.
Application code and source assets retain their respective licenses. This independent project does not imply endorsement by NASA, JPL, Google, MongoDB or other credited providers.