RAVEN

Radar Adaptive Vision Encoders — Live Demo
CVPR 2026 Demo Track
Anuvab Sen1, Mir Sayeed Mohammad1, Saibal Mukhopadhyay1
1Georgia Institute of Technology
0.27 GMACs sub-frame
compute
170× less compute
0.35 M model
parameters
≥50 FPS on edge
hardware

Live Demo

Attendees walk in front of the radar and see themselves detected in real time. The dashboard streams three views simultaneously: a camera feed with RAVEN's BEV detections overlaid, the radar-only detection output in a bird's-eye view, and a live energy-per-frame plot showing how the adaptive early-exit rule saves compute on simple scenes and ramps up only when needed.

RAVEN live demo dashboard with camera feed, BEV detection output, and energy-per-frame plot
RAVEN demo dashboard. (Left) Camera feed with detected people and AGVs labelled in real time. (Middle) Radar-only BEV detection output, showing range, angle, and class for each target. (Right) Live energy-per-frame plot — RAVEN exits early on simple scenes for up to 82 % energy savings, with a session-average of about 73 %.

Architecture

RAVEN processes the channel-wise sampled radar signal directly, without ever forming the full range–angle–Doppler cube. Per-receiver state-space encoders preserve antenna-level phase structure; a cross-antenna attention beamformer fuses them into a chirp-wise virtual array; and a slow-time chirp SSM produces the final detection and segmentation maps. The streaming design supports anytime inference — the model can emit a decision after any chirp prefix once the latent dynamics have saturated.

Animated RAVEN architecture: per-RX SSM encoders, attention beamformer, chirp-wise virtual array, and chirp SSM decoder
RAVEN architecture. Per-RX SSM modelers compress each receiver's fast-time samples into a compact token; the attention beamformer learns a steering-like fusion across antennas; the chirp-wise virtual array feeds a slow-time chirp SSM with calibrated early-exit decoding.

Hardware

Based on availability, the demo runs on either a LoCoBot WX250s mobile robot or a laptop, streaming raw ADC data from a TI IWR1843BOOST 77 GHz FMCW radar over a DCA1000EVM capture card. End-to-end inference happens on the edge — no cloud, no GPU cluster.

LoCoBot WX250s mobile robot platform with cameras and radar

LoCoBot WX250s

Mobile robot with 2D LiDAR, RGB-D and stereo cameras, plus an Intel i3 edge processor — the platform we deploy RAVEN on for live demos.

TI IWR1843BOOST FMCW radar evaluation board

TI IWR1843BOOST

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

Radar
TI IWR1843BOOST 76–81 GHz · 3 TX / 4 RX TDM-MIMO · 12 virtual antennas
Data Capture
DCA1000EVM 1 Gbps Ethernet real-time ADC streaming
Camera
Co-registered RGB Visual reference only — not used by the model

Results at a Glance

RAVEN delivers state-of-the-art detection and segmentation on both major automotive radar benchmarks while using orders of magnitude less compute than prior radar perception models — making it the first practical fit for always-on edge deployment.

Summary results poster showing RAVEN's accuracy versus compute trade-off across benchmarks