Data infrastructure startup Encord has closed on $60 million in new funding to help move robots, drones and other autonomous physical artificial intelligence systems from the lab into real-world production settings.

Today’s Series C round was led by Wellington Management and new investors Bright Pixel and Isomer Capital. Existing backers including Y Combinator, CRV, N47, Crane Venture Partners and Harpoon Ventures also piled back in, bringing the startup’s total amount raised to $110 million.

Encord, officially named Cord Technologies Inc., is the creator of an automated “data development platform” that helps companies to prepare the data needed to train physical AI models that power autonomous vehicles, robots and drones. Its software helps to address four key steps in creating AI training datasets, namely data management, data curation, model evaluation and data annotation.

It consolidates those tasks into one platform and automating them. In that way, it creates an audit trail that developers can use to dig into their physical AI models and understand why they generate specific outputs. In turn, that makes it easier to refine and improve their models and achieve better outcomes.

The round comes at a time when the physical AI industry is getting close to an inflection point, as thousands of robots and autonomous vehicles globally begin to come online, following years of intensive research and development and pilot projects. According to Encord, it’s anticipated that more than 400 million intelligent robots will come online in the next four years, pushing the physical AI industry to eclipse $30 billion annually.

Encord positions its data infrastructure as the perfect platform for handling the notoriously complex multimodal data streams that robots and autonomous systems generate, made up of audio, video, sensory data, 3D point clouds and other signals. Legacy infrastructure, the startup says, wasn’t designed to process these complex data types.

The challenge is that, unlike standard large language models that are trained mostly on the public internet, physical AI models require masses of proprietary, real-world data. Most of this comes from proprietary sources, including robot telemetry, sensor feeds and video footage that’s captured in the field. Such datasets tend to be massive, extremely diverse and very computationally intensive to process.

Encord co-founder and co-Chief Executive Ulrik Stig Hansen said there’s still far too much focus on model size, when the bottleneck isn’t really model muscle power, but data readiness. “You can have the most sophisticated model in the world, but it will still fail if the data feeding it is incomplete, inconsistent and misaligned with real-world conditions,” he said. “That’s the problem we solve.”

Since its last funding round around 18 months ago, Encord has grown significantly, with the data volume on its platform growing from just over 1 petabyte to more than 5 petabytes, representing three times the volume of information that was used by OpenAI Group PBC to train GPT-4. Over the same timeframe, its revenue has surged more than 10 times, the company said.

Encord works with more than 300 physical AI teams globally, serving clients including Toyota Motor Co.’s mobility subsidiary Woven, which develops software for autonomous cars, and the drone developers Zipline International Inc. and Skydio Inc. Those customers use its platform to support the entire model development lifecycle, spanning pretraining data generation to aligning models with human feedback after deployment.

Encord’s other co-founder and co-CEO Eric Landau said the company will use the money from today’s round to expand into new markets and accelerate its platform development. “The companies winning in physical AI understand what others are just beginning to realize,” he said. “It’s that the model is only as good as the data behind it. We’re building the infrastructure that makes data usable, not just once but continuously as these systems learn and improve.”

Image: Encord