AI Model Training Challenges Addressed by New Platform

Alex de Vigan has launched a new venture after his previous business was impacted by the rise of generative AI. The entrepreneur is now focused on Physicl, a platform that leverages technology initially developed for creating product renderings in e-commerce.

Physicl has secured $10 million in seed funding from Inovia Capital to help world model designers and robot builders overcome training challenges through more realistic simulations. This investment aims to utilize the tech stack built by de Vigan's previous company, nfinite, which was known for its 3D modeling capabilities.

The arrival of transformer-based AI models led nfinite to shift focus towards generative imagery. However, de Vigan saw an opportunity in re-purposing his team's expertise and technology to address the limitations of current robot training methods. Researchers have been struggling with the accuracy and cost-effectiveness of traditional approaches.

According to de Vigan, Physicl addresses a gap in AI capabilities: its ability to understand, interact, and manipulate space and objects. This is particularly relevant for researchers developing physical AI models to enable robots to handle complex tasks like manipulating small objects.

Ensuring that simulated worlds accurately reflect real-world physics has proven to be a significant challenge for researchers developing artificial intelligence models to enable robots to perform complex tasks.

Physicl's solution lies in leveraging the pipeline developed by its previous company, which transforms unstructured and incomplete real-world information into a high-fidelity physical representation with precise geometry, materials, and physics. This technology can now be applied to create more effective physical AI models for robots.

The startup is entering a crowded market where various companies are collecting, labeling, or generating data in different modalities to aid foundation model builders and robot operators. The space is still evolving, but Physicl's focus on post-training models on difficult edge cases sets it apart from other players.

Physicl has released Studio, a product featuring 10,000 3D environments and 100,000 assets representing various objects, allowing users to create scenes and generate synthetic data for training robots on specific problems. This can be particularly useful in situations where robots struggle with tasks like distinguishing between objects under different conditions, such as when an object is partially hidden or in shadow.

Physicl's collaboration with SpAItial has led to significant improvements in the fidelity of its products, including the newly released Echo-3 model.

SpAItial's tool generates 3D simulations from 2D imagery, and its CEO and founder Matthias Niessner explained that his company needed high-fidelity synthetic data to achieve this effectively. Physicl's platform provides such data, allowing SpAItial to train its models more accurately.

Physicl envisions itself not just as a data provider but also as a tool for ongoing synthetic data generation. This involves transforming unstructured information into high-fidelity 3D representations on an agentic workflow scale.

According to Niessner, the current process of curation is time-consuming and labor-intensive, involving manual review of large datasets. Physicl's platform aims to automate this process by generating data on the fly as needed for model training.

Physicl's tooling allows customers to feed in vast amounts of unstructured information, such as images or LIDAR capture data, and have it transformed into high-quality 3D representations that can be used to generate multiple permutations for training models.

Physicl's innovative tooling enables customers to input vast amounts of unstructured data, such as images or LIDAR capture information, and transform it into high-quality 3D representations that can be used to generate multiple applications.

These transformed representations can be leveraged for various purposes, including generating multiview reconstructions, 3D scene understanding, and even autonomous driving. The technology's versatility allows customers to tap into a wide range of possibilities, from improving product design to enhancing the capabilities of robots and self-driving vehicles.

By harnessing the power of Physicl's tooling, companies can unlock new avenues for innovation and growth, ultimately propelling their industries forward with more accurate and detailed 3D models.