4D Imaging Radar: Arbe Robotics Leads Industry Advancement in Direct Velocity Perception
Key Takeaways
- ▸4D imaging radar measures velocity directly in a single transmission burst, eliminating the frame-to-frame tracking problem that limits optical systems in dynamic environments
- ▸Arbe Robotics' 48x48 MIMO antenna array with 2,304 virtual channels achieves ~1° angular resolution—approximately 100x finer resolution than 2018-era automotive radar systems
- ▸Three cascaded FFT transforms extract range, Doppler velocity, and azimuth/elevation from raw radar data, producing a four-dimensional perception cube every 30 milliseconds
Summary
Traditional optical systems like stereo cameras can identify objects but cannot directly measure velocity without frame-to-frame comparison and tracking. 4D imaging radar solves this fundamental limitation by measuring range, azimuth, elevation, and radial velocity in a single burst of radio waves, making it critical for robots operating in dynamic environments with moving people, vehicles, and other robots.
4D imaging radar achieves this through MIMO (Multiple-Input-Multiple-Output) antenna arrays that multiply virtual receiver channels, combined with a three-step FFT (Fast Fourier Transform) process. The first FFT extracts range from frequency-modulated chirp signals, the second pulls Doppler velocity from phase progression across multiple chirps, and the third extracts angle from phase differences across virtual antennas. Arbe Robotics has advanced this technology with a 48x48 MIMO configuration generating 2,304 virtual channels, achieving approximately 1-degree angular resolution in both azimuth and elevation—roughly 100 times finer than automotive radar systems from 2018.
The industry is exploring different architectural approaches to 4D radar processing. While traditional automotive radars perform FFT computations on-chip and transmit sparse detection points, NVIDIA demonstrated a centralized processing architecture at GTC 2026 with Chinese radar maker ChengTech. This approach streams raw radar samples (approximately 540 megabytes per second for a five-radar vehicle setup) to the main computer for processing, representing a shift from sensor-centric to compute-centric radar perception.
- Industry is shifting toward centralized radar processing architectures that trade increased raw data bandwidth (540 MB/s) for processing flexibility and AI integration on main compute platforms
Editorial Opinion
4D imaging radar represents a fundamental capability leap for robot perception in dynamic environments. While Arbe Robotics has demonstrated the technical sophistication possible with advanced MIMO arrays, the industry trend toward centralized processing suggests that future competitive advantage will depend as much on software architecture and AI inference as on hardware specifications. As robots become increasingly prevalent in human-shared spaces, the ability to directly measure velocity—not infer it—could prove essential for safe, responsive autonomous systems.
