You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$25,000 left to reach the funding goal.
I want to make something that feels impossibly magical. I want a hyper-realistic avatar in a hyper-realistic world that forms as you speak. You should be able to enter a world you want, explore it as you please, and learn about anything.
The guide should understand what I am looking at, even as I move. It should be aware of the surroundings and able to transform the world as I ask over voice. I want a lifelike guide and natural voice in the same coherent world.
The more important thing is to discover or create a stack that can reliably make us do that. I want to know whether we are making real progress or getting trapped in incremental loops. That is where I want to look for a genuine engineering breakthrough.
This grant would test one part of that goal: keeping the world coherent as we move and change it. A scene can look convincing but have an unusable doorway. Objects can change when we return. The guide needs to see the same world we see. I request $25,000 to build and test an open component for this problem. A minimum of $5,000 would fund a smaller pilot.
Sciescape is my personal project. I am Navish Kumar, an PhD student at university of Basel. I'm using coding agents to help build it. The project has no university sponsorship.
The full experience is unfinished. Walking, a lifelike avatar, and natural voice still need to work together. I do not yet have evidence that others have adopted these tools.
Can one shared scene state keep these parts consistent? Rooms, openings, and objects will keep their identities. Rendering, movement, edits, and the guide will use the same state. The system will track the scene version and check edits before display. If an edit fails, the previous state will remain intact.
I propose 12 weeks to reproduce failures, test baselines in pilots, build the component, and compare the results.
Four conditions:
The current pipeline that uses depth panoramas.
Shared scene state.
Separate interaction state.
A public baseline that runs and has a compatible licence.
I want to test complete journeys through these scenes. Can we follow a route and pass through its openings? Does an object keep its identity when we return? Does an edit persist? Does the guide describe what is actually in its view? Separate evidence from the renderer will check those guide claims.
I will report visual quality, latency, memory, cost, and uncertainty for each scene. Timeouts and refusals will count as failures. An independent engineer will reproduce the test runner and review the scoring. I still need to appoint that person.
Success means tools that others can run and better results across complete journeys. I will also report changes in quality and cost. Text2Room, DreamScene360, and explicit scene representations already exist. I need to test whether this combination helps. If it does not, I will publish that result too.
The full $25,000 request covers:
$16,200 for my research and implementation: 180 hours at $90.
$4,800 for independent engineering and reproduction: 60 hours at $80.
$3,500 for GPU and host compute.
$500 for storage and transfer.
I propose 15 hours each week alongside my PhD. I will spend only after funding and terms are agreed.
The $5,000 pilot covers 40 of my hours ($3,600) and 10 hours of independent review ($800). It also covers compute ($500) and storage and transfer ($100). It would produce a small prototype and replay of failures. Six scenes, two conditions, and one seed give 12 exploratory journeys. This pilot would not fund the full study, a new avatar, or the complete voice experience.
Shared state may preserve faulty geometry or reduce speed and visual quality. Compute, model behaviour, and licences may limit the comparisons. I will test these risks early. If the idea fails, I will preserve useful replay tools. The study will use synthetic scenes and licensed voices, without private recordings or likenesses of real people.
A public viewer, synthetic scenes, and replay tools will work without the private Sciescape application. Replay of saved scenes will need no paid API. I will release new funded code and documentation under Apache-2.0. Examples and evaluation materials will use suitable open permissions, subject to rights in the source materials. I will exclude credentials, restricted assets, and model weights from other providers. The film will keep its separate noncommercial media licence.