Inverted AI

What Inverted AI does

Inverted AI develops leading, best-of-class, directable, human-like Non-Playable Characters (NPCs) for vehicle simulators, critical for achieving fully autonomous vehicles. Their proprietary technology, Imagining the Road Ahead (ITRA), produces realistic multi-modal predictions and directable, human-like NPCs from widely available data. This allows autonomous vehicle companies and others to save billions in data collection. Inverted AI's mission is to solve the infraction-free planning problem, utilizing a capital-efficient drone-based data pipeline to collect one of the biggest and most accurate behavior datasets globally. They offer software to transform and modernize AV/ADAS development and research, including API client libraries, open-source lightweight simulators, and RL environments.

Products

• ITRA™ PlannerA low-TOPS, data-driven L2++ neural driving policy and planner with post-perception object-list and bird’s-eye-view input, fully validated against hundreds of millions of driving scenarios. It is sensor-agnostic, RTOS-independent, and works with any map representation, delivering smooth, human-like driving behavior.
• ITRA™ VerifyA comprehensive suite of composable tools designed to verify and validate autonomous vehicle (AV) and ADAS system planners using a post-perception approach. It includes extensive scenario and log libraries, high-performance simulation, and reporting tools.
• INITIALIZEA core product that augments log-replay and procedural simulation trees with reactive, realistic, and diverse scenarios, used for initializing simulations by placing NPCs on a map.
• DRIVEA core product that augments log-replay and procedural simulation trees with reactive, realistic, and diverse scenarios, used for stepping simulations by driving NPCs.
• SCENARIOA core product that augments log-replay and procedural simulation trees with reactive, realistic, and diverse scenarios.
• Cloud APIScalable tools and enterprise-grade cloud APIs that dramatically improve simulation efforts and ADAS/AV development.
• Logs (Real Logs)A huge catalogue of high-quality naturalistic traffic and scenario data sourced from drones flown all over the world, facilitating analysis and comprehension of traffic participants' behavior and interactions.
• Logs (Synthetic Logs)In-house generated synthetic data using models trained on real-world information, producing high-fidelity simulations of traffic and scenarios. Available with selected HD maps or custom maps.
• Dynamic LogsCustomizable logs with real-time reactivity using APIs, allowing tailoring of agents to specific needs and instant generation of high-quality simulations.
• TorchDriveSimAn open-source lightweight 2D driving simulator serving as an environment for ADAS/AV controller development. It delivers high-performance object-level simulation.
• TorchDriveEnvAn efficient and simple reinforcement learning environment that allows users to train and evaluate their driving models with abundant features and customized scenarios.
• C++ SDKA detailed reference for the C++ library, providing access to key functions like BLAME, DRIVE, INITIALIZE, and LOCATION_INFO, and allowing direct access to the underlying REST API.
• Python SDKA detailed reference for the Python library, providing access to key functions like BLAME, DRIVE, INITIALIZE, LIGHT, LARGE_DRIVE, LARGE_INITIALIZE, LOCATION_INFO, and SIMULATION MANAGER, and allowing direct access to the underlying REST API.

Key capabilities

• Reactive, realistic, and behaviorally diverse NPCs
• Proprietary technology and deep generative models for human-like behaviors
• Scalable, accelerated development for AV/ADAS, autonomous robots, and smart cities
• Low-TOPS, data-driven L2++ neural driving policy and planner
• Comprehensive suite of tools for AV/ADAS system verification and validation
• Post-perception approach for planner validation
• Cloud APIs for simulation and development
• Augmentation of log-replay and procedural simulation with reactive scenarios
• Compatibility with any real-world or simulated location
• Huge catalogue of high-quality real and synthetic logs
• API-based access for dynamic synthetic data generation and customization
• Open-source lightweight simulators (TorchDriveSim)
• Efficient reinforcement learning environments (TorchDriveEnv)
• C++ and Python SDKs for API integration
• Fully-validated low-TOPS realtime prediction and planning modules
• Sensor-agnostic and RTOS-independent planner
• Robust and reliable driving with smooth, human-like behavior
• Extensive scenario and log libraries (over 150M concrete scenarios on 3,000 HD maps across 30 countries)
• High-performance object-level simulation
• Integration with Weights and Biases for reporting
• Cloud-based scenario editor
• Built-in perception sensor noise models
• Automatic blame attribution
• Data-driven NPC models

Use cases

• Accelerated development of safe technology for autonomous vehicles (AV)
• Accelerated development of advanced driver assistance systems (ADAS)
• Development of autonomous robots
• Smart city simulations
• Training improved AV 2.0 end-to-end networks (AV 3.0 approach)
• Integrating prediction and planning modules into onboard stacks (AV 1.0)
• Validating post-perception stacks
• AV/ADAS controller development
• Training and evaluating driving models in reinforcement learning environments
• Analysis and comprehension of traffic participants' behavior and interactions
• Developing, testing, or validating ADAS and AD systems
• Conducting studies on traffic, driver, or pedestrian behavior
• Infrastructure monitoring and planning
• Generating cost-effective driving solutions to difficult and uncommon scenarios for AV 2.0 developers
• Comprehensive expected and worst-case planner testing

AI approach

Inverted AI develops proprietary deep generative models and machine learning algorithms to create human-like NPCs for vehicle simulators, autonomous vehicles, and ADAS. They also provide data solutions including real and synthetic logs, and tools for verifying and validating AV/ADAS system planners. Their approach includes AV 3.0 for data augmentation and end-to-end network training, and AV 1.0 for prediction and planning modules.

Tech named: deep generative models, machine learning, neural driving policy, prediction algorithms, planning algorithms, probabilistic programming, amortized inference, reinforcement learning, PyTorch

Industries served

• Autonomous Vehicles
• Automotive (ADAS)
• Robotics
• Smart Cities
• Software Development

What it says sets it apart

• Leading, best-of-class, directable, human-like NPCs
• Proprietary technology and deep generative models for realistic, human-like behaviors
• Capital-efficient drone-based data pipeline for large and accurate behavior datasets
• Proprietary, world-leading prediction and patent-pending correct-by-construction planning algorithms
• APIs called over half a billion times
• Fully-validated low-TOPS realtime prediction and planning modules
• AV 3.0 approach for data augmentation and improving AV 2.0 end-to-end networks
• Ability to work with any real-world or simulated location
• Comprehensive suite of tools for verification and validation, decoupling planner from perception system
• Extensive scenario and log libraries (150M+ concrete scenarios, 3,000 HD maps, 30 countries)
• High-performance simulation (22 hours driving per GPU hour on RTX 3080)
• PyTorch-native simulator with DDP support and Hydra configuration management
• Cloud-based scenario editor and dynamic synthetic data generation via API

Funding rounds we track

From the AI funding tracker — rounds as reported by the linked publications.