Zeroframe

What do they actually do

Zeroframe builds and sells curated, real‑world video and 3D datasets for robotics, 3D vision, and foundation “world” models. The catalog includes projects like Urban3D (multiview videos with 3D reconstructions) and EgoExplore (long-duration egocentric video with synchronized sensors) packaged for research and product teams https://zeroframe.ai/urban3d https://zeroframe.ai/egoexplore https://zeroframe.ai/.

They capture head‑mounted and 360° recordings, sometimes with IMU/LiDAR, then run structure‑from‑motion (e.g., COLMAP) to produce calibrated assets such as camera poses, sparse point clouds, depth/NeRF‑ready transforms, and annotations (captions, scene/weather/time labels, IMU streams) https://zeroframe.ai/urban3d https://zeroframe.ai/egoexplore.

Distribution is consultative today: sample assets are public (e.g., Urban3D sample and previews), while full datasets are delivered on request via email or a booking link. There’s no self‑serve pricing or storefront yet, and pages note “FULL DATASET (IN PROGRESS)” for some releases https://zeroframe.ai/urban3d https://zeroframe.ai/egoexplore https://zeroframe.ai/.

Who are their target customer(s)

• Academic or industry research labs building perception/world models: They need diverse, real‑world multimodal recordings with calibrated outputs and rich annotations, but collecting and preprocessing this data is slow and expensive. Zeroframe offers curated datasets targeted at this use case https://zeroframe.ai/ https://zeroframe.ai/egoexplore.
• Robotics product teams working on embodied agents or manipulation: They require long‑duration egocentric footage, synchronized IMU/finger‑tracking, and labeled manipulation examples; producing aligned streams and interaction labels is brittle and costly. EgoExplore/EgoDex target these modalities https://zeroframe.ai/egoexplore.
• 3D / neural‑rendering and AR/VR teams: They need many high‑resolution multiview captures plus camera geometry and reconstruction outputs that plug into NeRF/SfM pipelines; scaling capture and reconstruction takes specialized tooling and compute. Urban3D provides NeRF‑ready assets https://zeroframe.ai/urban3d.
• Foundation‑model training organizations: They want large, diverse, legally clear datasets delivered in training‑ready formats, but acquiring/licensing and preparing bulk data is bespoke and slow. Zeroframe frames datasets as inputs for world models and distributes full sets via consultation today https://zeroframe.ai/.
• Domain‑specific product teams (retail, indoor mapping, sports/emotion analytics): They need vertical collections that reflect real operating environments, while public datasets rarely match. Zeroframe lists domain projects like Grocery_Explore and Indoor_Space to address these gaps https://zeroframe.ai/.

How would they acquire their first 10, 50, and 100 customers

• First 10: Run tightly scoped pilots with top labs and robotics teams via YC and direct outreach, offering free or discounted full‑dataset access for feedback, benchmarks, and case studies; use the existing consultative contact/Cal.com flow to deliver https://zeroframe.ai/urban3d https://zeroframe.ai/egoexplore https://www.ycombinator.com/companies/zeroframe.
• First 50: Publish dataset papers/preprints and larger sample packs (e.g., on public mirrors) and host workshops/challenges at relevant conferences to drive citations; leverage early case studies and light outbound to close labs and mid‑size robotics teams https://zeroframe.ai/urban3d.
• First 100: Launch a simple self‑serve catalog with standardized licensing tiers and cloud delivery, then run targeted enterprise outreach (foundation‑model orgs, robotics vendors) with pilot‑to‑paid paths and options for subscriptions or custom collections https://zeroframe.ai/ https://zeroframe.ai/egoexplore.

What is the rough total addressable market

Top-down context:

Published estimates put AI training datasets around ~$2.6B in 2024 (with some sources up to ~$3.2B), robotics benchmark/datasets at ~$1.37B, and 3D scanning at ~$4.3B (of which only a slice maps to dataset sales) https://www.grandviewresearch.com/industry-analysis/ai-training-dataset-market https://www.marketsandmarkets.com/Market-Reports/ai-training-dataset-market-153819655.html https://dataintelo.com/report/robotics-benchmark-datasets-market https://www.grandviewresearch.com/industry-analysis/3d-scanning-industry.

Bottom-up calculation:

Core TAM ≈ AI training datasets (2.6–3.2B) + robotics datasets (1.37B) + 10–30% of 3D scanning as dataset/content (0.4–1.3B) ≈ $4.4–$5.9B (2024 baseline) https://www.grandviewresearch.com/industry-analysis/ai-training-dataset-market https://dataintelo.com/report/robotics-benchmark-datasets-market https://www.grandviewresearch.com/industry-analysis/3d-scanning-industry.

Assumptions:

• We include only image/video/multimodal portions of the AI training‑dataset market relevant to perception/world models.
• We add the robotics dataset/benchmark segment once (to avoid double‑counting general AI dataset figures).
• We attribute just 10–30% of the 3D scanning market to curated dataset/content sales (excluding hardware and most software/services).

Who are some of their notable competitors

• Scale AI: Enterprise data engine and collection/labeling services (including multi‑sensor and autonomy use cases); competes for bespoke real‑world dataset programs and delivery.
• Synthesis AI: Synthetic data platform for vision and 3D human/scene datasets; an alternative to purchasing large real‑world captures for training.
• Parallel Domain: Generates synthetic 3D worlds and sensor data for autonomy/robotics; substitutes for real‑world multiview and sensor captures in some workflows.
• Waymo Open Dataset: Large open autonomous‑driving datasets with LiDAR/camera; widely used baseline/alternative for driving and perception research instead of bespoke purchases.
• Ego4D: Massive open egocentric video dataset (Meta and partners); an alternative source for long‑duration egocentric footage used in embodied/world‑model research.