You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
Robot learning has a data problem that nobody measures in the open. Labs collect huge amounts of first-person operator video, and the tooling that says whether any of it is good training data lives inside those labs, closed. I am building the open version of that tooling.
Two pieces are already shipped and public:
1) video-benchmark - a Python CLI plus a WebGPU web app built around a simple question: is this operator video usable? It scores footage with classical vision metrics, temporal checks, audio analysis, and hand-visibility signals, and the browser app runs everything locally on your own machine. Code at github.com/shubhxho, live demo at video-benchmark-lemon.vercel.app.
2) ego-world-model-v1 - a public dataset on Hugging Face with per-frame depth, hand and object masks, action labels with confidence and reason, and IMU streams, so other people can build against the same reference: huggingface.co/datasets/shubhxho/ego-world-model-v1.
With this funding I will build three things: (a) portable dataset-quality reports - one command, runs locally, works on any team's own footage; (b) an open provenance registry, where shared training data carries a public record of where it came from and what was done to it, so anyone can audit it; and (c) open reference benchmarks, so dataset quality can be compared across the field instead of argued about.
Why this is public-benefit: evaluation and provenance are audit infrastructure. When anyone can run them locally, a lab, a researcher, or a regulator can inspect what went into a robot's training data instead of trusting a closed pipeline - including catching low-quality, duplicated, or poisoned data. Everything I produce is open source (MIT/Apache-2.0 dual-licensed) and documented in public.
Milestones (target $50,000 over 4 months; if only the $10,000 minimum is raised, I start at that level with scope trimmed to milestone 1):
Milestone 1 (months 1-2, $20,000): portable dataset-quality reports - a one-command, locally runnable quality report for any first-person robotics video dataset: scoring, distributions, failure slices.
Milestone 2 (months 2-3, $15,000): open provenance registry - a public registry where datasets register origin, transformations, and licensing, plus reference tooling for reading and writing records.
Milestone 3 (months 3-4, $15,000): open reference benchmarks and release - benchmark suites comparing dataset quality, docs, packaged releases.
Budget across the total: about 70% compute (training and evaluation runs), 20% data collection, 10% release and documentation. I will post public devlog updates on this page at each milestone.
Funds go only to the project described here. I am applying as an individual (Shubh Gupta), not an organization. No equity, no tokens, no investment terms - this is a non-dilutive grant request.
Just me. I am Shubh Gupta, 18, an independent open-source developer in Bengaluru, India, working full-time and in public. I received an Emergent Ventures grant for this work.
On this exact problem: I designed and shipped video-benchmark (github.com/shubhxho, demo at video-benchmark-lemon.vercel.app) and published the ego-world-model-v1 dataset on Hugging Face. Other public work: based-chess-rs, a no_std Rust chess engine whose evaluation network is distilled from its own search, and nanoharness, a terminal harness for coding agents - both on my GitHub.
The most likely failure is adoption: teams keep whatever scoring they already use, and the registry stays empty. The mitigation is to make trying the tools cost nothing - one command, local, no signup. A technical failure mode is that the quality signals that predict training outcomes on my footage do not transfer to other teams' setups; the reference benchmarks exist to measure that transfer instead of assuming it. If the project fails outright, the code, the dataset, and the benchmarks stay public and usable, and unspent funds are returned.
An Emergent Ventures grant for this work. Nothing else.