The hardest problem for an AI that needs to move through the world isn't the model architecture. It's the data. Most training sets for physical tasks are synthetic, stitched together from game engines or static 3D scans, missing the messy, continuous physics of real life. Volumes, a New York-based startup, is building its business on the premise that this gap represents a fundamental bottleneck. The company operates a fleet of multi-sensor volumetric capture rigs, turning real-world events into dense, licensable spatiotemporal datasets it calls "ground-truth reality" [Volumes, retrieved 2024]. For AI labs training robots, autonomous vehicles, or digital humans, this is the raw material they can't easily generate themselves.
The data wedge
Volumes' product suite is built around a hierarchy of fidelity. At the base is raw, unprocessed volumetric capture from its sensor arrays. The next layer is processed reconstructions: static 3D scenes using Gaussian Splatting, a newer technique that can produce highly detailed point-cloud-like renders more efficiently than traditional meshes. The flagship offering is 4D Gaussian Splatting, which adds the temporal dimension, capturing how objects and people move and interact over time [Volumes, retrieved 2024]. Finally, the company uses this captured data to generate synthetic volumetric scenes, expanding training sets with physically plausible variations. The throughline is measurability; the output isn't just a video, but a precise, queryable copy of space, time, and physics.
Why the market is moving its way
The timing hinges on two converging trends. First, the AI industry's pivot toward "embodied" or "physical" intelligence, where models must reason about and act in three-dimensional environments. The research labs pushing this frontier are increasingly data-constrained. Second, the recent maturation of Gaussian Splatting as a practical 3D representation. It offers a compelling trade-off between visual quality and computational efficiency, making it more viable for large-scale dataset creation than older methods. Volumes is positioning itself at the intersection, aiming to be the supplier of record for this nascent but critical input.
The competitive landscape
Volumes operates in a specialized niche. Its direct competitors include companies like Evercoast and 4DViews, which also offer enterprise volumetric video capture and Gaussian Splatting services. The differentiation Volumes emphasizes is its focus on AI training data as the primary product, rather than volumetric video for media or entertainment. However, the competitive set is broader than it appears. Any provider of synthetic data for robotics, from NVIDIA's Omniverse to countless simulation startups, is addressing a similar customer need with a fundamentally different approach. The Volumes bet is that the authenticity of real-world capture will outperform the scalability of simulation for certain high-stakes applications.
The technical breakdown centers on the 4D Gaussian Splatting pipeline. Capturing a dynamic scene requires synchronized data from multiple sensors (likely RGB cameras, depth sensors, and possibly LiDAR). The system must then fuse these streams into a coherent 4D model, a computationally intensive process that involves estimating the position, scale, rotation, and opacity of millions of anisotropic 3D Gaussians across hundreds of frames. The output is a compact file that can be rendered from any viewpoint and, critically, queried for physical properties like occlusion, contact forces, and trajectory.
Scaling this operation presents clear challenges. The capture process is physical and logistical, requiring trained crews to deploy sensor rigs on location. The computational cost of processing petabyte-scale raw captures into clean 4D splats is non-trivial. And the business model, selling high-value datasets to a relatively small number of well-funded AI labs, must prove it can achieve the volume needed to support the underlying infrastructure costs. The company's undisclosed seed round [Crunchbase, retrieved 2026] will be tested against these capital-intensive realities.
Sources
- [Volumes, retrieved 2024] Volumes - Capturing Reality | https://www.volumes.cloud/
- [Crunchbase, retrieved 2026] Seed Round - Ellis | https://www.crunchbase.com/funding_round/joinellis-seed--475e0042
- [evercoast.com, retrieved 2026] Evercoast | Multicamera, 4D Spatial Video | Enterprise Volumetric Video | Gaussian Splatting | https://evercoast.com/