CVPR 2026 Demo Track

RAVEN

Radar Adaptive Vision Encoders

Real-time human detection from raw FMCW radar signals on edge hardware. Physics-inspired streaming architecture with adaptive early-exit inference.

0
GMACs
0
Less Compute
0
Parameters
0
FPS on Edge
01 — Live Demo
Dashboard
Attendees walk in front of the radar and see themselves detected in real time. Camera feed, BEV detection output, and adaptive energy plot all running on edge hardware.
RAVEN Demo
02 — Architecture
Physics-Inspired Streaming
Per-receiver SSMs, cross-antenna attention beamformer, and chirp-wise SSM backbone with anytime early-exit inference.
RAVEN Architecture
03 — Hardware
Edge Deployment
Based on availability, the demo runs on a LoCoBot mobile robot or a laptop, streaming raw ADC data from a TI IWR1443BOOST radar sensor.
LoCoBot

LoCoBot WX250s

Mobile robot with radar, 2D LiDAR, RGB-D and stereo cameras, Intel i3 edge processor.

TI IWR1443BOOST

TI IWR1443BOOST

76–81 GHz FMCW radar-on-chip. 3TX, 4RX, 12 virtual elements in TDM-MIMO. Up to 4 GHz chirp bandwidth.

Radar
TI IWR1443BOOST, 76–81 GHz, 3TX/4RX TDM-MIMO, 12 virtual antennas
Data Capture
DCA1000EVM, 1 Gbps Ethernet real-time ADC streaming
Camera
Co-registered RGB, visual reference only
Compute
Intel i3 edge processor (LoCoBot) · NVIDIA RTX-class laptop GPU
Model
0.35M params · ≤1.02 GMACs full-frame · ≤0.27 GMACs sub-frame
Speed
≥50 FPS with sub-frame latency via adaptive early exit
04 — Results
Performance at a Glance
State-of-the-art detection and segmentation accuracy at orders of magnitude less compute than prior radar perception models.
Results