Cheaper and safer LIDAR sensors using AI

State-of-the-art performance

Inference time of 5ms (including pre-processing and post-processing) on NVIDIA GTX 1080. Over 98% accurate within 0.01 meters with ground-truth data. We are able to improve range by 10 - 20 meters, and enhance object detection and classification task accuracy by 2X

Easy to use

We provide a highly stable runtime piece in form of a ROS node, and can train and finetune the model to your hardware and environment. We are compatible with multiple software and hardware stacks and LIDAR sensors

Extensive training and validation

Our model is trained on over 1 MN+ LIDAR sweeps from complex environments collected from 8, 16, 64 and 128 channel LIDAR sensors. We have set benchmarks on depth prediction, estimation on the KITTI Vision benchmarking suites (unpublished)'s technology can be used to improve perception in adverse conditions as well as boost the object detection and classification capabilities from high resolution sensors

Reduce the cost of LIDAR sensors by over 90%

A single sweep from a Velodyne VLP16 sensor upsampled by 4X

Want to set up a demo? Get in touch!