Inverted

Inverted AI is a Vancouver-based company developing generative artificial intelligence models that produce non-playable characters (NPCs) for driving simulation. The NPCs are designed to be reactive, realistic and behaviourally diverse in the manner of human road users, and are intended to support scalable development and testing of autonomous vehicles and advanced driver assistance systems (AV/ADAS), autonomous robots, smart cities and simulations more broadly. The company states its mission is to solve the infraction-free planning problem.

The company delivers its NPC engine as a cloud service accessed through a REST API, with Python and C++ SDKs layered on top; the Python client SDK is open source and distributed via PyPI. Integration is designed as synchronous co-simulation between a customer's local simulator and the NPC engine running on Inverted AI servers, with support for HD maps (including CARLA town maps), traffic-light state, waypoint management, region-based agent initialization, scene plotting and simulation logging. Product components described publicly include Drive (reactive human-like NPC driving behaviour), Initialize (realistic and diverse agent placement), plus Blame (attribution of causative agents and reasons in simulated collisions) and Scenario (generation of complete test scenes populated with pedestrians, cyclists, cars, buses and traffic lights).

Inverted AI frames its offering across three market segments it labels AV 1.0, AV 2.0 and AV 3.0. AV 1.0 addresses modular AV/ADAS architectures, for which the company offers low-TOPS realtime prediction and planning modules for onboard integration and post-perception validation and verification. AV 2.0 refers to end-to-end learned driving stacks. AV 3.0 is described as a data augmentation scheme for training improved end-to-end networks, in which Inverted AI drives all vehicles including the ego vehicle on traffic-light-controlled HD maps, with demonstration videos rendered using NVIDIA COSMOS.

Founding story

Inverted AI was co-founded by Frank Wood, who serves as CEO and is an associate professor of computer science at the University of British Columbia and a Canada CIFAR AI Chair at Mila, and Adam Ścibior, CTO and an adjunct professor of computer science at UBC who previously completed a postdoc with Wood at UBC. Wood's earlier career includes academic positions at Oxford and Columbia, a postdoc at UCL's Gatsby Computational Neuroscience Unit, a PhD from Brown University, and entrepreneurial experience running and selling the content-based image retrieval company ToFish! to Time Warner and serving as CEO of Interfolio.

Business model

Inverted AI sells access to its NPC simulation engine as an API-based service. Usage requires an access key obtained by creating an account on the company's user portal; new users receive keys preloaded with an API access budget, and users affiliated with academic institutions generally receive credits sufficient to conduct research at no cost. The company also positions prediction and planning modules for integration into customers' onboard AV stacks and for post-perception validation and verification.

Access to the Drive, Initialize and related APIs is gated by API keys tied to an access budget, with free credits for new and academic users, indicating a usage-metered commercial API model alongside engagements to integrate onboard prediction and planning modules.

Traction

The company reports that its APIs have been called over half a billion times, and that its prediction and planning algorithms have been validated in hundreds of millions of real-life scenarios across thousands of locations in tens of countries. Its Drive and Initialize products were in market as of the seed announcement, with Blame and Scenario in beta testing. The open-source Python SDK repository shows 1,670 commits and 17 stars. The company has raised institutional funding from Yaletown Partners and other investors.

Latest developments

The company's applications materials describe AV 3.0, a data augmentation scheme for training end-to-end AV networks in which all vehicles including the ego are driven on traffic-light-controlled HD maps, with demonstrations rendered using NVIDIA COSMOS. It also offers realtime, low-TOPS prediction and planning modules for onboard integration in modular AV 1.0 stacks and for post-perception validation and verification. The careers page lists no open positions.