The Promise of Physical AI: Bridging the Simulation-Reality Gap
The promise of physical AI is that engineers will be able to program physical agents in the same way they do digital agents.
In today’s rapidly evolving technological landscape, physical AI stands at the frontier, offering a transformative potential to replicate the programmability of digital agents in physical agents. However, we’re not there yet. Robotics, a key player in physical AI, faces significant challenges due to the lack of data from physical spaces. To train machines effectively, companies often resort to building mock warehouses for testing, while an industry has emerged around monitoring factory lines and gig workers to train deep learning models for robotic operation.
The Role of Simulation in Overcoming Data Limitations
An alternative to physical testing is simulation—detailed virtual replicas of real-world environments that could provide the data and workspaces roboticists need to conduct their work in a scalable manner. Antioch, a startup specializing in simulation tools for robot developers, aims to bridge what the industry calls the simulation-reality gap: the challenge of making virtual environments realistic enough for robots trained within them to operate reliably in the physical world.
“How can we do the best possible job of closing that gap, so that the simulation looks exactly like the real world from the perspective of your autonomous system?” asked Harry Mellsop, co-founder of Antioch.
To pursue this goal, Antioch has successfully raised an $8.5 million seed round, valuing the company at $60 million. The round was led by venture capital firm A* and Category Ventures, with additional participation from MaC Venture Capital, Abstract, Box Group, and Icehouse Ventures.
Antioch’s Journey and Vision
Founded in May last year by Mellsop and four co-founders, Antioch is based in New York. Among the founders, Alex Langshur and Michael Calvey previously co-founded Transpose, a security and intelligence startup, and sold it to Chainalysis. The other two co-founders—Collin Schlager and Colton Swingle—bring experience from Meta Reality Labs and Google DeepMind, respectively.
The need for better simulation is crucial for many large autonomy companies. In the self-driving car domain, for instance, Waymo uses Google’s global DeepMind model to test and evaluate its driving model. This approach aims to minimize data collection requirements, a significant cost in autonomous vehicle technology development.
Empowering Startups with Simulation Tools
Building and using these models to test robots requires a different skill set than creating a self-driving car. Antioch aims to provide a platform to solve this problem for new companies lacking the capital to develop it independently. These smaller companies often cannot afford to build physical testing arenas or drive sensor-equipped cars over millions of miles.
“The vast majority of the industry doesn’t use simulation at all, and I think we now understand clearly that we need to move faster,” Mellsop emphasized.
Antioch’s product is compared to Cursor, the AI-based software development tool. It enables robot builders to create multiple digital instances of their hardware and connect them to simulated sensors that replicate real-world data inputs. These environments facilitate testing edge cases, performing reinforcement learning, and generating new training data.
Ensuring Realistic Simulations
Achieving realistic simulations is crucial, as the physics of the simulation must align with reality. Antioch begins with models from Nvidia, World Labs, and others, developing domain-specific libraries for ease of use. By working with multiple customers, Antioch gains a depth of context to refine its simulations, unmatched by any single physical AI company.
“What happened with software engineering and LLMs is just starting to happen with physical AI,” said Çağla Kaymaz, a partner at Category Ventures. “The challenges are different. In the physical world, the stakes are much higher.”
Antioch’s Focus and Future Aspirations
Currently, Antioch focuses on sensor and perception systems, crucial for automated vehicles, agricultural and construction machinery, and aerial drones. While aspirations for physical AI to power generalized robots replicating human tasks are more distant, Antioch’s initial engagements include large multinationals already investing heavily in robotics.
Adrian Macneil, an executive at self-driving startup Cruise and founder of Foxglove, supports Antioch as an angel investor. “Simulation is really important when you’re trying to build a safety case or handle very high-precision tasks,” he stated at the Ride AI conference in San Francisco. “It’s not possible to travel enough miles in the real world.”
Macneil envisions tools akin to those driving the SaaS revolution—platforms like GitHub, Stripe, and Twilio—emerging to support physical AI. “We need a lot more of the full toolchain to be commercially available,” he noted.
“We all honestly believe that anyone building an autonomous system for the real world will be doing it primarily in software form in two to three years,” Mellsop concluded. “This is the first time that autonomous agents can iterate on a physical autonomy system and close the feedback loop.”
Real-World Applications and the Path Forward
Experiments in this direction are already underway. David Mayo, a researcher at MIT’s Computer Science and Artificial Intelligence Laboratory, uses the Antioch platform to evaluate LLMs. In one experiment, Mayo tasks AI models with designing robots and uses the Antioch simulator to test them, even pitting them against each other in simulated competitions.
However, to usher in an era of AI engineers, there’s ongoing work to bridge the digital-real world gap. Achieving this would enable developers to create a data flywheel, crucial for the success of category leaders like Waymo, where engineers can confidently expect consistent model improvements.
Companies looking to replicate this success must develop these tools themselves or acquire them.
For more on Antioch’s journey and the evolving landscape of physical AI, visit the original source Here.
“`
