應用研究
TIER IV Open-Sources METEOR: One Network Uses Eight Cameras for Simultaneous Perception and Path Planning on Jetson Orin
The 54M-parameter METEOR builds a BEV representation from eight camera feeds and outputs 12 driving tasks in a single forward pass. Its INT8 version runs in approximately 67–70 ms on Jetson AGX Orin. The code, ONNX model, weights, and demo data are public, but the deployment version disables temporal memory, and no public autonomous-driving benchmark is yet available for cross-model comparison.
Japanese autonomous-driving company TIER IV has released METEOR, a multitask end-to-end model intended for L2++ research. It takes input from eight surround-view cameras, along with camera calibration data and vehicle speed, without relying on high-definition maps. After a ResNet-34/FPN extracts image features, depth-guided inverse perspective mapping projects them into an 800×500 BEV feature map with a resolution of 0.2 meters per cell. Multiple task heads then output lane segmentation, depth estimation, 2D/3D detection, occupancy and flow, traffic-signal recognition, risk fields, predictions for other road users, and three ego-vehicle trajectories with confidence scores. Rule-based safety checks run separately outside the neural network.
The released model has just 54M parameters. Its convolutional backbone uses 2:4 structured sparsity, while the planning branch remains dense. The team provides a PyTorch checkpoint, an opset 17 ONNX model, TensorRT Python/C++ runtimes, and six anonymized demo scenarios. According to official measurements, inference on one eight-camera frame using INT8, CUDA Graphs, and zero-copy input on Jetson AGX Orin has a median latency of approximately 67–70 ms. The model card also notes that the TensorRT engine must be rebuilt locally for the target GPU architecture and software version.
Another experimental focus of the release is the development workflow. All training labels were generated by the CoMET teacher ensemble from camera, LiDAR, and localization data. Agents then handled coding, training, failure analysis, quantization, hardware profiling, and rollback between candidate versions. Humans remained responsible for setting objectives and reviewing results, so “zero human-written code” does not mean an absence of human decision-making or safety responsibility. LiDAR is used to generate training supervision, while the standard deployment inputs remain camera-only.
The repository also documents a significant failure: a data-caching bug caused end-to-end training to receive all-zero historical features for an extended period. After genuine historical features were restored, trajectory performance from the existing weights deteriorated by 33%–38%. As a result, the current ONNX model removes temporal memory entirely, making it effectively a single-frame model and leaving a gap between the deployed model and the temporal-fusion capability shown in the architecture diagram. The project does not yet provide results on public benchmarks, closed-loop simulation, or safety-case evaluations, and the released demo data has no ground-truth labels. For now, engineers should regard METEOR as a reproducible edge-deployment baseline rather than a driving system ready for direct on-road use.