TIER IV Open-Sources Five Autonomous Driving Models for Autoware

TIER IV released five end-to-end autonomous-driving AI models as open source on November 11, 2026, spanning camera-only systems and configurations that add LiDAR, radar and high-definition maps. The lineup — NVIDIA Alpamayo 2 Edge, AutoE2E, METEOR, OnePlanner and LatentDrive — gives automakers and technology partners several options for optimizing models and integrating them into vehicle systems. All five connect environmental perception with driving-trajectory generation, but they differ sharply in sensor requirements, planning approach and readiness for the road.

The release matters beyond TIER IV itself. Open-source autonomy stacks compete with proprietary systems from Waymo, Tesla and larger suppliers, and TIER IV is betting that manufacturers want escape hatches from single-vendor lock-in. The models arrive as concrete outcomes for Autoware 2.0, the next-generation architecture TIER IV detailed on November 2, 2026, which combines AI-based and conventional rule-based approaches, adds a flexible indexing scheme for modular software integration and provides a Reference Design framework intended to ease commercial adoption across vehicles, applications and operating environments.

What are the five models?

The models range across three orders of magnitude in size, and none is finished.

• NVIDIA Alpamayo 2 Edge — 2 billion parameters, versus 34 billion in Alpamayo 2 Super. It uses only camera images and reasoning to generate driving trajectories. Autoware integration is complete; TIER IV is still optimizing the model and adapting it to real vehicles.
• AutoE2E — 79.9 million parameters, developed jointly by the Autoware Foundation and TIER IV. It runs on computing platforms beyond NVIDIA hardware and supports configurations from cameras alone to systems with LiDAR and radar.
• METEOR — 16 million parameters, developed independently by TIER IV and aimed primarily at privately owned passenger vehicles. It runs on NVIDIA and Renesas semiconductor platforms.
• OnePlanner — combines perception and motion planning, generates six-second trajectories, and has complete Autoware integration.
• LatentDrive — uses self-supervised learning from camera video, produces four-second trajectories, and remains in integration and in-vehicle validation.

In short, only two of the five models — Alpamayo 2 Edge and OnePlanner — have finished Autoware integration, and even those still lack completed in-vehicle work.

How do the architectures differ?

AutoE2E uses a fast–slow architecture that pairs a Reactive Model for real-time decisions with World and Reasoning Models that predict environmental changes and support driving decisions. That split mirrors how many robotics systems separate reflexive control from slower deliberation.

METEOR takes a different route. Its convolutional neural network (a network type that processes grid-like image data) uses a bird's-eye-view representation to handle perception through trajectory generation in one network, without relying on HD maps. It accepts images from one to eight cameras and can add LiDAR and radar.

OnePlanner currently uses BEVFusion to create a bird's-eye-view representation from LiDAR point clouds, and passes features extracted during perception directly to the planner. TIER IV says this reduces intermediate conversions into object positions and velocities. It also incorporates HD maps, route information, traffic-signal states and vehicle motion states.

LatentDrive is the most experimental of the five. It uses V-JEPA, a self-supervised learning approach, to represent features from camera video in a latent space — a compressed mathematical representation — and generates trajectories without intermediate object detection and tracking. It needs no HD map and supports one or three cameras.

Each model trades something away: parameter count, sensor payload or integration maturity. Buyers will need to match those trade-offs to their vehicle platforms.

What does AutoE2E v1.0 actually deliver?

The Autoware Foundation's Robotaxi Working Group, sponsored by TIER IV, released AutoE2E v1.0 on November 11, 2026, under the Apache 2.0 license for research and commercial use, with trained checkpoints also available.

The release includes the Reactive Model, which runs at 10 Hz, fuses six camera views into a bird's-eye representation and plans 6.4 seconds of acceleration and steering. The world-model and reasoning layers are still being developed to work with the same planner — meaning the headline v1.0 release is currently a partial system.

The foundation reports a specific benchmark: the fine-tuned AutoE2E model achieved an average displacement error of 0.744 meters over three seconds in recorded-data validation, ranking second among eight open-weight configurations in its comparison. That number deserves scrutiny. In that open-loop test, models predicted from logged camera frames and were scored against the human driver's actual path, and inputs varied across models — AutoE2E received a map raster and route while other models did not necessarily get the same inputs.

The result therefore concerns trajectory prediction on recorded data, not driverless road performance. The gap between the two is the central caveat in the entire announcement.

AutoE2E is integrated with Autoware for open-loop evaluation, and real-vehicle closed-loop testing is underway to assess interactive driving. The foundation says deployment performance and resource requirements remain subject to platform-specific evaluation, with further on-road trials planned for 2027.

The project targets applications from L2++ hands-free driving assistance (driver supervision still required) to future L4 driverless mobility. L4 remains a development goal rather than a current capability.

How does TIER IV plan to commercialize this?

For METEOR, TIER IV uses CoMET, the automated labeling platform within Co-MLOps, together with agentic AI to automate data labeling and model training, evaluation and improvement — eliminating, the company says, the need for manual coding and labeling. Automakers and component suppliers can then optimize the open-source model for their own vehicle specifications and performance requirements.

Founder and CEO Shinpei Kato framed the multi-model strategy as an anti-lock-in play. It is intended, he said, to help manufacturers and suppliers "internalize technologies according to their applications and performance requirements, without depending on a particular AI model or semiconductor platform."

NVIDIA automotive vice president Ali Kani offered a complementary pitch for the largest model in the set, saying the Alpamayo 2 Edge integration gives developers an open foundation for moving the technology from research into their own vehicles. NVIDIA has a commercial interest here: Alpamayo 2 Edge targets NVIDIA hardware, even as AutoE2E and METEOR deliberately support other silicon.

TIER IV's stated next steps are further performance improvements, in-vehicle optimization, commercialization with automakers and partners through Autoware, and expansion across vehicle types and applications.

What is still unsaid?

The announcement does not claim any vehicle currently driving driverless with these models. It does not publish closed-loop test results, safety-case documentation or regulatory engagement. The 0.744-meter displacement figure — the only hard benchmark disclosed — comes from an open-loop comparison with non-uniform inputs, which limits its comparability.

Autoware 2.0 supplies the shared architecture, but AutoE2E's unfinished closed-loop testing shows that open-source availability and validated vehicle deployment remain separate milestones. Until the 2027 on-road trials report results, the five models are a credible research and integration toolkit, not a proven commercial autonomy product.