TIER IV open-sourced an end-to-end driving model it said an agentic AI developed

TIER IV published a reference end-to-end driving model called METEOR as open source on September 8, releasing the source code, the trained model and the training recipe together. The company said the model takes video from eight surround cameras and needs no high-definition map, producing road-surface segmentation, three-dimensional object detection with heading and velocity, an occupancy grid, traffic-light state and an ego trajectory from a single network. A separate rule-based component checks the trajectory the network proposes. The company said the target is Level 2++ driving.

The model runs on one NVIDIA Jetson Orin, converted through ONNX to TensorRT and quantized to eight-bit integers with per-layer sensitivity analysis. Training labels came entirely from CoMET, TIER IV’s automated labeling platform, which the company described as an ensemble of twelve large models. The training data came from Co-MLOps, its data-sharing platform in Japan, collected with a rig carrying four lidars and eight cameras; lidar was used only to generate labels, and the deployed model reads cameras alone. TIER IV said it used NVIDIA Cosmos to transfer recorded scenes into rain, night and snow conditions and to generate rare cases.

The claim TIER IV placed at the center of the release concerns how the work was done. It said an agentic AI carried out the coding, the training and the in-vehicle optimization, running one experimental variable per round against acceptance criteria fixed in advance, rolling changes back automatically when they failed, mining failures for the next round, and keeping a record of how much accuracy each change bought per millisecond of runtime.

The company said the model is scheduled to be demonstrated on an in-vehicle computer at Automotive World 2026 in greater Tokyo, which runs from September 9 to 11. Members of Co-MLOps may fine-tune and extend the released model.