Seeing Distance Without Radar Tesla’s Machine Learning System That Estimates Object Properties From Camera Images Alone

Tesla has filed a patent for a machine learning system that can estimate the distance and properties of objects from visual camera images alone – without relying on dedicated distance-sensing hardware such as radar or lidar. The invention is central to Tesla’s vision-only approach to autonomous driving and represents a significant step toward making accurate depth perception achievable through cameras and trained neural networks, potentially reducing the cost and complexity of autonomous vehicle sensor systems.

The Problem

Autonomous driving systems must perceive the world around the vehicle in three dimensions – knowing not just what objects are present, but how far away they are, how fast they are moving and what path they are likely to take. Traditional autonomous vehicle sensor suites rely on a combination of cameras, radar and lidar (light detection and ranging) systems to build this three-dimensional picture.

Lidar systems, in particular, generate highly accurate point-cloud distance data but are expensive, mechanically complex and add significant bulk to a vehicle. Radar provides distance and velocity data but with limited resolution. Cameras provide rich visual information but are inherently 2D – the image captured by a single camera does not directly encode depth information the way a distance sensor does.

The challenge Tesla and others have pursued is whether a neural network can learn to infer depth and other three-dimensional object properties from camera images alone – by learning the visual cues that correlate with distance in the real world. If this is achievable at sufficient accuracy and reliability, it opens the door to autonomous driving systems that rely entirely on cameras, potentially at substantially lower cost than sensor-fusion approaches.

What This Invention Does

The system consists of one or more processors trained to receive image data from a camera mounted on a vehicle and use that image data as input to a trained machine learning model. The model is trained to identify the distance of an object from the vehicle based on visual image data alone.

The key innovation in the training approach is that the model is trained using pairs of data: a training image from the camera alongside the output of an “emitting distance sensor” (such as a radar or lidar) capturing the same scene at the same moment. By correlating what the camera sees with what the distance sensor measures, the model learns the visual features that predict distance – essentially learning to replicate the distance sensor’s measurements using only what the camera sees.

Once trained, the deployed model can estimate object distance and properties from camera images without requiring the distance sensor at all – the sensor is only needed during training, not during operation. This approach allows Tesla to develop and validate vision-based depth perception by leveraging existing sensor data during development, then deploy a streamlined camera-only system in production vehicles.

Key Features

Vision-only depth estimation. The trained model estimates object distance from camera images alone during operation, without requiring radar or lidar – reducing sensor complexity and potentially lowering vehicle cost significantly.

Sensor-supervised training. The model is trained using correlations between camera images and concurrent distance sensor outputs, enabling it to learn accurate depth inference from labelled real-world data without manual annotation.

Machine learning model architecture. The system uses a trained machine learning model – enabling the depth estimation capability to improve with additional training data and to generalise across diverse real-world conditions.

Vehicle-integrated deployment. The system is designed for integration into a vehicle’s onboard computing system, processing camera feeds in real time to support autonomous navigation and driver assistance functions.

Scalable data collection. Because training data is collected by vehicles already equipped with both cameras and distance sensors during normal operation, the training dataset can be expanded continuously as more vehicles accumulate driving miles.

Who Is Behind It?

Tesla, Inc. is the Palo Alto-based electric vehicle and clean energy company, known for its Autopilot and Full Self-Driving driver assistance systems. The inventors are James Anthony Musk, Swupnil Kumar Sahai and Ashok Kumar Elluswamy. This application is a divisional of AU 2020224581, which originated from PCT/US2020/017290 (published as WO 2020/171983), claiming priority to US Patent Application No. 16/279,657 filed 19 February 2019. The application is managed by Spruson & Ferguson in Sydney.

Why It Matters

The race to develop reliable, affordable autonomous driving technology is one of the defining technological contests of the current era. Tesla’s bet on a camera-only sensor approach – rejecting the lidar systems used by many competitors – has been controversial in the industry but is backed by a clear commercial logic: cameras are far cheaper than lidar, and if machine learning can replicate lidar-quality depth perception from camera data, the cost advantage of a vision-only system could be decisive for mass-market deployment.

The machine learning approach described in this patent – training on sensor-correlated data and then deploying without the sensor – is a compelling demonstration of how AI can effectively transfer the capability of expensive hardware into a learned model. With the IPC classification covering visual scene understanding (G06V 20/00), the patent is directly relevant to the core perception technology at the heart of Tesla’s autonomous driving development programme.

---

AU 2026201529 was published in the Australian Official Journal of Patents on 19 March 2026 and is open for public inspection. Patent applications represent inventions that are sought to be protected and do not necessarily reflect commercially available products.

Related Concepts

Lidar has been the dominant technology for distance sensing in autonomous vehicles, producing highly accurate point-cloud maps of surroundings. Tesla’s contrarian bet on computer vision alone – using machine learning models trained on correlated camera and sensor data – aims to replicate lidar-quality depth perception at camera cost. If proven reliable, it could make autonomous driving technology significantly cheaper and more scalable for mass-market vehicles.