Setup, navigation, source frames, legacy kinematic recordings, service configuration and API reference. For physical episodes, use the simulation guide.
Use Node.js 24 and npm. The 3D view needs a WebGL-capable browser with hardware acceleration. Public imagery and the three surface sites need no credentials. Globe images and the Moon/Earth site packages are bundled locally; Dingo Gap tiles still require network access.
git clone https://github.com/k2jac9/rocket-academy-first-contact.git
cd rocket-academy-first-contact
npm ci
npm run devnpm run dev starts both Vite and the Node API on loopback by default. Open the Vite URL, normally http://127.0.0.1:5173. The API listens at http://127.0.0.1:8787; Vite proxies /api to it. Both services are reachable only from the local machine with these defaults.
A missing .env.local is valid for exploration and the kinematic mission tools. Gemini planning and Atlas persistence require the optional server configuration below. Unsaved kinematic episodes live in page memory. Robot lab requires Atlas and the isolated Python environment described in simulation/README.md.
The explorer, catalog, lab, replay and dialogs use one shared interface theme: local Inter/Workbench/IBM Plex Mono fonts, white surfaces, pastel image panels and blue actions. Navigation preserves mission and episode state. See the design system for the shared tokens and component rules.
Open Robot catalog, or go directly to http://127.0.0.1:5173/#/catalog/humanoid or #/catalog/quadruped. The Omblet-based catalog includes five models in each of six classifications: humanoids, quadrupeds, mobile manipulators, companions, drones and field robots. All 30 profiles include editorial images and dated technical specifications (531 fields in the November 14, 2026 snapshot). Household robots and standalone arms are excluded from browsing, search and comparison.
Search by model, maker or purpose; filter by manufacturer, availability, starting price or simulation support. Select up to four robots to compare, show only differing technical fields, copy a comparison URL, or export a CSV with source links and dates. Unknown values stay explicit, and specifications with different measurement conditions remain separate. The shortlist survives reloads locally. The catalog uses local images/fonts and does not need Atlas or Gemini.
Catalog coverage is separate from simulation support. Jackal links into the episode lab; Go2 has 3D inspection only. The older Husky model in the lab is not the catalog's Husky A300. The other catalog entries currently provide specifications and images, not runnable robot models. Source records and editorial portraits are attributed to Omblet; see NOTICE.md.
Open Scenario studio (#/studio) to prepare a mission brief and world prompt, import an optional reference image and exploration video, and retain their provenance in Atlas. The Google Genie adapter is versioned and records Genie 3, Genie 4 or an unknown model separately from an optional exact backend revision. There is no live Genie generation or automated robot-control connection. The adapter reports unverified programmatic access; the existing Gemini planner key does not establish Genie entitlement.
- Write the mission intent and world prompt, then save the brief to Atlas directly in Scenario studio.
- Optionally import a PNG/JPEG/WebP reference (up to 8 MiB) and an MP4/WebM recording (up to 32 MiB). Original bytes are stored in GridFS with server-computed SHA-256 hashes. A brief without a video is explicitly identified as such.
- Select a saved exploration and freeze a scenario version. New versions can target Moon / Apollo 17, Earth / Amboy Crater, or Mars / Dingo Gap. Historical design-only versions remain immutable.
- Open that version in Robot lab, freeze a measured terrain patch and run. Robot models, duration, speed and seed remain locked until the scenario is detached. The server rejects settings that differ from the saved version.
- The physical episode records the scenario hash and
inspiredBylink in Atlas, local recovery output and MCAP. The scenario freezes bounded pilot settings; it does not compile an arbitrary natural-language objective into a cooperative policy.
Imported generative experiences default to exploration/review only. Their reference images are visual conditioning, not measured collision geometry. No action timeline, physical ground truth or training eligibility is inferred from video. See the integration contract and the platform implementation map.
- Install the local simulation environment.
- In Explore worlds, choose Apollo 17 or Amboy Crater for a prepared starter location, or select Dingo Gap and place a rover inside the reconstruction.
- Select Use patch in robot lab to freeze a complete 16 m terrain crop. Missing elevation cells stop capture.
- Choose the model for each robot, duration, speed and seed. Select Generate & save episode.
- Replay the stored states, inspect the camera previews, and download the compressed MCAP. Use Episode Explorer at the top of Robot lab to filter saved episodes by world, terrain site, robot model, end reason and UTC date range, then reopen or download a matching episode.
These are experimental contact-physics episodes: source robot visual/collision geometry and inertias, rigid wheels, assumed controls/friction and simple virtual sensor rendering. Detailed appearance does not establish physical accuracy. Model revisions and asset hashes are recorded; keep the bundled mesh assets alongside MCAP exports. The terrain sample interval is not a measurement-accuracy claim. See asset attribution and the architecture and remaining milestones.
- Open Mars, the Moon, or Earth with the NASA globe imagery layer.
- Open Mission tools, then select Load 20 m example, or choose Place rover and Set goal to select nearby surface points. The example provides valid coordinates even when the current browser cannot render WebGL.
- Select Baseline plan. Review the initial observation, drive, and final observation commands. The baseline drives at 0.75 m/s and does not call a generative model.
- Select Run locally. The executor creates a bounded sequence of recorded states; it does not send motion commands to a physical robot.
- Replay the recorded episode and inspect the rover position and trajectory. Playback speed affects viewing time, not the recorded timestamps.
- Select Export MCAP to download the episode. Export plan JSON separately downloads the planned mission manifest.
- With the optional services configured, choose Gemini plan for a model-generated proposal, or Run & save to Atlas to execute the validated plan on the server and store its run and recording.
The example is a 20 m arc on the application's declared reference sphere. It starts at native (radius, 0, 0) and moves toward native +Y along the equator. It does not identify a surveyed landing site or prove that a 20 m stretch of the displayed imagery is traversable. The overview images are much coarser than rover-scale detail.
Changing the source, layer, markers, or objective invalidates the corresponding plan and episode. A Gemini failure is reported explicitly; the application does not silently relabel a baseline plan as AI. Run & save to Atlas creates a server episode from the mission and plan, rather than uploading a previously recorded browser episode. Its run identifier and start time belong to that saved execution.
Keyboard controls operate when the canvas has focus. Tab to it or click inside the view. An active placement tool shows a center reticle.
| Action | Control |
|---|---|
| Orbit or navigate | Drag; arrow keys |
| Zoom | Mouse wheel; + / − |
| Pan across Dingo Gap | Shift + arrow keys |
| Move faster around a globe | Shift + arrow keys |
| Place a marker with a pointer | Choose a placement tool, then click the visible surface |
| Place a marker with a keyboard | Move the surface under the reticle, then press Enter or Space |
| Reset to the complete globe or local site | Home, or the reset control |
| Cancel placement | Select the active placement tool again, or its cancel control |
Imagery-layer picks use an exact mathematical sphere. Terrain-layer picks use loaded terrain geometry. Clicking empty space does not create a marker. If an image fails to load, the view explicitly reports Reference sphere only; the plain sphere can still support a kinematic example. No measured terrain is invented to replace a failed tile request.
| View | Data | Surface model | Access |
|---|---|---|---|
| Mars imagery | NASA Mars texture, 1440 × 720; Viking imagery processed at USGS | Reference sphere, radius 3,396,190 m; no elevation | Bundled, unaltered image with SHA-256 |
| Moon imagery | NASA LRO CGI Moon Kit, 2048 × 1024 | Reference sphere, radius 1,737,400 m; no elevation | Bundled, unaltered image with SHA-256 |
| Earth imagery | NASA Blue Marble Next Generation, July 2004, 5400 × 2700 | Reference sphere, radius 6,371,000 m; no elevation | Bundled, unaltered image with SHA-256 |
| Dingo Gap | NASA/JPL Curiosity colorized ground reconstruction, release URS297920 |
Streamed local 3D tiles on Mars | Public upstream tiles |
| Apollo 17 | LROC NAC stereo DTM and orthophoto | 512 m lunar site, 2 m elevation posts | Bundled, hashed package |
| Amboy Crater | USGS 3DEP 1 m DEM and USDA NAIP imagery | 512 m Mojave site | Bundled, hashed package |
| Moon terrain | Cesium ion asset 2684829 |
Provider's streamed planetary terrain | Restricted ion browser token |
| Mars terrain | Cesium ion asset 3644333 |
Provider's streamed planetary terrain | Restricted ion browser token |
A bounded Earth terrain site is available at Amboy Crater; a global Earth terrain layer is not implemented. Blue Marble is a static image, not live weather or a WGS84 ellipsoid. The lunar color map is processed for visual presentation; it is not the accompanying LOLA displacement map. Imagery detail, terrain geometry, and authoritative collision geometry are separate inputs.
The source catalog is the application authority for URLs, radii, frame descriptions, and supported source/layer combinations. NOTICE.md records asset links and credits, including the bundled Moon file's checksum. The upstream descriptions are NASA Mars Image Texture, NASA CGI Moon Kit, and NASA Blue Marble base maps.
Reference globes use right-handed, body-centered XYZ in metres, with +Z toward the north pole. The model convention is:
x = radius × cos(latitude) × cos(longitude)
y = radius × cos(latitude) × sin(longitude)
z = radius × sin(latitude)
Moon and Earth imagery have approximate spherical registration. Mars texture longitude registration is not established by this application. These are declared modeling frames, not a claim of surveyed geodesy. The display applies an X rotation of −π/2, and picks are transformed back to native source coordinates before storage. The sphere's radius is enforced by the mission validator.
Dingo Gap uses the native local XYZ frame of the pinned NASA dataset. Its revision is 76444abbb4faff1b05f46a09b81ca2c24602b333. The display applies an X rotation of +π/2, mapping source XYZ to (x, -z, y); source +Z points downward in that view. No geographic origin, north/east alignment, or latitude/longitude conversion is asserted for the local reconstruction.
Optional ion terrain retains provider-defined body-fixed coordinates. Asset IDs identify hosted datasets, not immutable snapshots, so exports mark these sources provider-managed-unpinned. The Mars texture is pinned to a NASA repository commit; the bundled Moon image is pinned by its bytes; the Earth image URL is not an immutable revision. Streamed tile level of detail is view dependent and is never recorded as a fixed collision mesh.
Mission JSON uses rocket-academy.mission schema 2.0.0. It records the source/layer, surface model, radius, imagery provenance, native frame, units, marker positions, and explicit unvalidated simulation flags. It excludes credentials and unrelated application state. The validator rejects unsupported fields, mismatched provenance, invalid coordinates, and inconsistent distances. The current UI has no mission-import workflow; historical 1.0.0 exports are not silently migrated.
The mission's distance card and dashed connection use Euclidean chord distance. Globe execution follows a shortest spherical arc with interpolated endpoint radii; Dingo execution uses a straight line in local source coordinates. Neither legacy kinematic execution samples terrain contact, slope, obstacles, or hazards, including when a measured terrain layer is visible. The separate MuJoCo Robot lab path uses a frozen heightfield; its contract is documented in the simulation guide.
Both planners produce the same versioned plan: a rationale and a sequence containing only drive_to, wait, observe, and stop. Shared validation runs before execution in both the browser and server. Generated code, arbitrary tools, and external action URLs are not accepted commands.
| Current execution bound | Value |
|---|---|
| Fixed simulation step | 0.2 s |
| Drive speed | 0.01–2 m/s |
| Maximum individual drive leg | 1,000 m |
| Maximum planned duration | 1,800 s |
| Maximum recorded frames | 10,000 |
| Maximum plan commands | 32 |
| Goal tolerance | 0.05 m |
The engine is first-contact-kinematic/1.0.0. It records planner kind, optional model/objective, commands, command events, source and local poses, observation labels, timestamps, and terminal result. A seed is retained for forward compatibility and explicitly marked unused-deterministic; the current executor has no stochastic dynamics. An observe command records derived kinematic state with synthetic: true, not a camera image or a physical measurement.
MCAP exports use indexed, uncompressed chunks, UTF-8 JSON messages, and embedded JSON schemas. The robot is represented by a generic geometric symbol. Official Foxglove schemas describe poses, frame transforms, and the robot scene; application schemas describe the mission, commands, state, observations, and result.
| MCAP topic | Contents |
|---|---|
/mission/run |
Mission/source identity, plan, planner, engine, seed semantics, and replay-frame metadata |
/mission/commands |
Command start/completion events |
/tf |
terrain_source → mission_local → robot transforms |
/robots/rover-1/pose |
Foxglove pose in the small mission-local frame |
/robots/rover-1/state |
Native and local state, speed, goal distance, and command index |
/robots/rover-1/observations |
Explicitly synthetic kinematic observations |
/scene/robot |
Foxglove scene representation of the rover symbol |
/mission/result |
Terminal outcome and remaining goal distance |
The mission-local replay frame is derived from the start and goal, with X forward, Y left, and Z up. Its explicit origin and quaternion connect it to the native source frame. MCAP timestamps use the recorded UTC start plus simulation elapsed time. Files can be inspected with compatible MCAP/Foxglove readers; the app does not yet import arbitrary MCAP recordings.
Every legacy kinematic run records physicsValidated: false, terrainContactValidated: false, imageSensors: false, and trainingRun: false. MCAP establishes a recording and replay path; it does not by itself make the data suitable for robot training.
Copy the example once, fill in the integrations you want, and restart the development command:
cp .env.example .env.local| Variable | Where it is used | Purpose |
|---|---|---|
MONGODB_URI |
Node server only | Atlas database-user connection string |
MONGODB_DB |
Node server only | Dedicated application database; the example uses first_contact, while an unset value defaults to rocket_academy |
GEMINI_API_KEY |
Node server only | Gemini API credential for planning |
GEMINI_MODEL |
Node server only | Optional model override; /api/status reports the configured/default model |
ELEVENLABS_API_KEY |
Node server only | Optional speech generation; requires Text to Speech and Voices read access |
ELEVENLABS_MODEL |
Node server only | Narration model; defaults to eleven_multilingual_v2 |
ELEVENLABS_VOICE_ID |
Node server only | Optional preferred voice from the returned account voices |
VITE_CESIUM_ION_TOKEN |
Browser | Optional public, restricted ion token for Moon/Mars terrain |
FIRST_CONTACT_API_PORT |
Node and Vite configuration | Optional API port; default 8787 |
FIRST_CONTACT_ALLOWED_ORIGINS |
Node server only | Optional comma-separated exact HTTP(S) origins |
Keep database credentials and Gemini keys server-only: never prefix them with VITE_. Node loads .env.local through its native environment-file support. Vite exposes configured VITE_ values to the browser and embeds those used by a production build. Do not put secrets in an objective, mission name, or other exported user text.
Use an existing cluster and a dedicated database user with readWrite access limited to the application database, including the collections and indexes created there. Restrict that user to the intended cluster where appropriate. An Atlas account user and a database user are different identities; see Atlas database-user configuration.
In the project's network access list, allow the current public egress IP of the machine running the Node service, preferably as a temporary entry during development. Use an exact IP or a narrowly scoped network; do not use an all-IP rule as the default setup. Set MONGODB_URI to the database user's connection string and MONGODB_DB to the database granted to that user. This project does not provision a cluster, purchase a tier, or create account credentials.
Atlas stores the following central records:
| Collection or bucket | Responsibility |
|---|---|
terrain_sources |
Versioned source/layer identities, body, coordinate metadata, and provider references |
robots |
Robot catalog revision and execution limits |
missions |
Validated mission manifests |
plans |
Validated plans and planner provenance linked to a mission |
runs |
Run state, complete episode, mission/plan references, and artifact metadata |
decision_events |
Execution commands, planner rationale/provenance, and terminal result |
GridFS run_artifacts |
MCAP bytes in run_artifacts.files and run_artifacts.chunks |
Run artifacts record byte length and SHA-256. Catalog records are upserted on explicit saves/runs; opening the catalog does not seed those records. Completed runs can be listed by body and their MCAP files downloaded. Failed or interrupted saves are not presented as completed runs. Same-request retries use idempotency keys; conflicting reuse is rejected.
Native coordinates remain attached to their body, source, and frame. The service does not apply Earth-specific geospatial distance calculations to Moon/Mars positions. Atlas vector retrieval, semantic failure search, and external object storage are future integrations.
Set GEMINI_API_KEY from your authorized Gemini account and, if necessary, GEMINI_MODEL to a model available to it. The service calls Google's structured-output REST interface with the validated mission and objective, then validates the returned plan against the same limits as a baseline plan. It does not give the model tools to execute code or drive a robot. The request sets store: false; provider account terms still govern API use.
Configured means the server has a key. It does not prove model access or quota. A successful planning request establishes that the selected account/model worked for that request. Authentication, quota, timeout, malformed output, and rejected-plan errors remain visible. The baseline planner stays a separately selected local option.
The public imagery globes need no ion token. For the Measured terrain layer on Moon or Mars, create a read-only public browser token restricted to asset 2684829 and/or 3644333 and to the application origins where it will be used. Account access and the provider's content terms apply.
Enter the token in connection settings for the current page session, or set VITE_CESIUM_ION_TOKEN before starting/building the frontend. Session entry overrides the configured token and is not written to browser storage or mission exports. A build-time token is visible in the delivered JavaScript; use only a token intended for public clients. Removing it from a delivered build requires rebuilding.
curl http://127.0.0.1:8787/api/statusThe response reports Atlas configured and connected separately, Gemini configuration/model, and execution limits. Each status refresh probes authenticated primary access and reports lastCheckedAt, latencyMs, and a safe error code when unavailable. With no optional credentials, both integrations report unconfigured; credentials remain outside the repository.
Run npm run check:atlas for read-only connection, index, and saved-artifact hash checks. Run npm run check:atlas -- --write to also test an isolated metadata/GridFS round trip through a fresh client and remove its diagnostic records afterward. The configured local Atlas connection passed these checks on November 14, 2026. See Atlas connection checks and recovery for evidence, behavior, and scope.
The API accepts local hosts and an explicit origin allowlist. Defaults cover localhost and 127.0.0.1 at the development, preview, and API ports. If you deliberately change the frontend origin, configure its exact origin and restart. This is a local-development boundary, not a substitute for authentication and authorization on a public service.
Database schema activation is intentionally deferred. New saves carry a storage-envelope version; the proposed validators passed isolated writer tests but are not installed on the live collections. See the schema-validation evidence, permission requirement and rollback commands before resuming Step 3 activation.
GET /api/episodes supplies Robot lab's Episode Explorer. It returns { episodes, nextCursor }, with metadata only and a default page size of 20. Existing consumers can continue reading the episodes array. Follow nextCursor to reach older recordings; a null cursor means the last page. The same filters apply to GET /api/episodes/stats, whose counts cover all matching complete physics episodes independently of page size. Neither endpoint returns sensor payloads.
| Parameter | Accepted values |
|---|---|
body |
earth, moon, mars, unknown |
site |
dingo-gap, apollo-17, amboy-crater, unknown |
robot |
husky, jackal, legacy scout / surveyor, unknown; matches any participating model |
termination |
duration_reached, rollover, terrain_boundary, unknown |
from, to |
Inclusive UTC calendar days, YYYY-MM-DD; missing dates are excluded when date filtering |
limit |
Integer 1–50, default 20; list endpoint only |
cursor |
Opaque cursor returned by the preceding list page; reuse the same filters; list endpoint only |
Omit a parameter to select all values. Unknown categories include missing or unsupported metadata. Unsupported/duplicate parameters, invalid calendar dates, reversed ranges and mismatched cursors return 422. Example: /api/episodes?body=earth&robot=husky&termination=rollover&limit=5; use the same query without limit for /api/episodes/stats.
Ordering is descending (startedAt, episode ID), with missing/non-string dates last. Refresh returns to the first page and includes newly saved episodes. This is a live library, not a frozen dataset: list and overview are separate reads, and edits/deletions can change membership. Completed records are not edited by this workflow. The cursor is a pagination marker, not an authorization token. See Step 2 verification and query-plan limits and the implementation plan.
Optional narration uses GET /api/speech/status, GET /api/speech/voices, and POST /api/speech with { text, voiceId }. The POST returns MP3 bytes, a SHA-256 header and cache-hit status. See ElevenLabs setup and limits.
The legacy mission/planning/run POST routes below use JSON with Content-Type: application/json and a 256 KiB body limit. Episode and exploration routes have their own contracts and limits in the simulation guide and Genie integration guide. Errors use { "error": { "code", "message", "retryable" } } without returning provider credentials.
| Method and path | Request / behavior |
|---|---|
GET /api/status |
Service version, integration status, model, and limits; no external credentials required |
GET /api/catalog |
Supported source variants and the current robot catalog |
POST /api/plans |
{ mission, objective }; requests and validates a Gemini plan; requires Gemini configuration |
POST /api/missions |
{ mission }; validates and saves a mission in Atlas |
POST /api/runs |
{ mission, plan, planner, seed? }; validates, executes on the server, writes MCAP to GridFS, and saves the run in Atlas |
GET /api/runs?body=mars&limit=20 |
Completed Atlas runs; optional body is earth, moon, or mars; limit is 1–50, default 20 |
GET /api/runs/:id |
A saved run and artifact metadata; id is a UUID |
GET /api/runs/:id/mcap |
The saved MCAP with Content-Length and X-Artifact-SHA256 headers |
Use a UUID Idempotency-Key on mission and run POSTs, and retain it when retrying the same operation. Omitting it causes a new UUID to be generated. Reusing it with different content returns a conflict. Plan generation is not an idempotent save and does not automatically retry provider calls.
Server saves additionally bound run JSON to 8 MiB and MCAP to 16 MiB; the browser writer has a 64 MiB memory cap. The service permits at most two concurrent plan requests and two concurrent run-save operations. History/detail/download and save routes require a usable Atlas connection. There is no server fallback that pretends a local record was stored remotely.
Episode Explorer also includes a native Atlas semantic-search prototype over factual summaries. See the semantic-search operator guide for activation, bounded reconciliation, provider costs and readiness checks. It uses Atlas-managed Voyage embeddings and does not require another Gemini key.
| Area | Responsibility |
|---|---|
| App and components | Planet/layer selection, mission editing, planning controls, replay, downloads, and connection state |
| Mission contract | Source catalog, coordinates, schema 2.0.0, validation, and provenance |
| Example mission | A 20 m reference-sphere arc |
| Terrain renderer | Three.js spheres, NASA tile streaming, native picking, robot/trail display, and resource cleanup |
| Episode engine | Plan validation and bounded deterministic kinematics |
| MCAP recording | Indexed writing, explicit schemas, timestamps, and frame transforms |
| Local API | Shared validation, route/origin limits, and integration errors |
| Gemini adapter | Bounded structured planning requests and output validation |
| Atlas adapter | Catalog/mission/plan/run records, GridFS, hashes, and save retries |
The current service uses the existing Node 24/TypeScript runtime and shared validators. The production frontend build is not a bundled server; Node executes the server TypeScript directly. The physical episode worker uses Python and MuJoCo; its environment and sensor contracts are documented in the simulation guide.
npm test
npm run typecheck
npm run buildThe tests exercise mission/plan constraints, globe transforms and picking math, keyboard navigation, credential scoping, startup deadlines, deterministic episodes, MCAP readback, and controlled API/provider failures. GitHub Actions runs installation, tests, and the production build. Consult the latest CI and QA output for actual results; this README does not substitute a frozen test count for current verification.
To serve the production frontend locally, run npm run preview. Start npm run start:api in another terminal if you need Gemini or Atlas; preview proxies /api but does not launch the service. npm run dev:ui and npm run dev:api are also available separately.
Verification is environment-specific. Current visual QA records globe/site rendering, robot inspection and episode replay. The Atlas check records authenticated reads/writes, direct GridFS hash verification and connection recovery from the configured local environment. Historical reports describe earlier slices and their limitations. Gemini generation and optional ion terrain still require their own authorized accounts and successful provider requests.
| Symptom | Next step |
|---|---|
| WebGL cannot start | Enable hardware acceleration or use a WebGL-capable browser; the example and local recording path can still be used |
| Plain globe with an imagery warning | Retry the image; check access to its host; no elevation is present in either case |
| Dingo Gap fails to load | Check raw.githubusercontent.com access and retry terrain |
| Moon/Mars terrain requests a token | Switch to public imagery or provide the restricted ion token |
| A plan exceeds the leg/duration limit | Use the 20 m example or choose closer waypoints |
| Local service is offline | Start npm run dev or npm run dev:api; local runs and MCAP downloads remain available |
| Atlas configured but disconnected | Check the database user, database name, current egress IP, and cluster network access |
| Gemini configured but planning fails | Inspect the reported model-access, quota, deadline, or schema error |
| API rejects an origin | Use a default local origin or explicitly configure the intended origin |
Dated evidence is kept in the QA reports. Check the version and environment recorded by each report; a successful check of an earlier slice is not evidence of successful live service calls in this one.
The current workflow connects measured terrain, source robot assets, shared-world MuJoCo execution, replay, MCAP and Atlas. Moon / Apollo 17 and Earth / Amboy Crater packages complement Mars / Dingo Gap. Scenario Studio records exploration provenance and freezes runnable pilot settings.
- Calibrated physical behavior. Validate suspension, contact, actuation, material and environment assumptions against relevant evidence. Current source geometry/inertias do not establish planetary qualification.
- Richer sensor and learning contracts. Add calibrated noise/latency, the required modalities, complete action timing, task rewards, dataset admission and frozen evaluation suites.
- More embodiments and policies. Extend beyond the runnable Husky/Jackal models; Go2 currently supports inspection. Robot catalog inclusion is not execution support.
- Verified generative interfaces. Keep Genie prompt/import workflows available while programmatic access is assessed. Cosmos observation generation and Isaac Sim/Isaac Lab workers remain proposed adapters.
- Governed datasets and learning. Add evidence-backed semantic retrieval, immutable dataset membership, training workers, evaluation and a measured curriculum loop.
See the platform design for architecture and acceptance gates.
Application code is licensed under Apache-2.0. Dependencies retain their upstream licenses. Scientific imagery, reconstruction products, and hosted services have their own source notices and terms; the application's code license does not relicense them.
The Mars, Moon and Earth overview JPEGs are bundled unaltered with provenance and SHA-256 checksums. Terrain tiles load from their upstream hosts. See NOTICE.md for exact sources, credits, and usage references. This is an independent prototype and does not imply endorsement by NASA, JPL, Cesium, Google, MongoDB, or the contributing data institutions.