01 — Sensor fusion 02 — Cinematic AV 03 — Rover + ROS 04 — Architecture 05 — Live dashboard

BEV · EVS · SNN — three views, one pipeline

Bird’s-eye object tracking, raw event camera stream and SNN spike raster — all from the same neuromorphic fusion model running on BrainChip Akida hardware.

BEV + EVS + SNN pipeline — Akida target BEV view
Bird’s-eye view
CONF —
94%
Energy reduction
2.1ms
Fusion latency
847Hz
Update rate
%
TCP confidence
Ready
Select a view above then press play. BEV shows fused object tracking with TCP confidence bars. EVS shows raw asynchronous event camera stream. SNN shows layer-by-layer spike raster.
00:0000:3001:05
Akida hardware target · ready

Four-panel fusion — EVS · BEV · SNN · TCP

All four sensor fusion layers running simultaneously. The long cut includes architecture breakdown and GPU benchmark comparison.

Phasor · AV fusion — four-panel view T+00:00
EVS — event stream
BEV — object track
SNN — spike raster
TCP — confidence
94%
Energy reduction
2.1ms
Latency
847Hz
Update rate
%
TCP confidence
Ready
Four-panel simultaneous view of the full AV sensor fusion pipeline. Select cut duration above then press play.
Architecture
EVS cam
Delta enc
SNN L1
LiDAR
Rate enc
SNN L2
IMU
Temporal enc
TCP fusion
Benchmark vs GPU baseline
Phasor
2.1ms
GPU
38ms
00:0000:3001:00
Ready · Akida hardware target

Full-stack robotics — perception to manipulation

Rover with Doosan M1013 6-DOF arm. ROS2 topic graph live. TCP confidence propagating from EVS camera through SNN fusion to arm controller — the complete closed-loop neuromorphic stack.

Phasor · Rover sim v0.4 — Doosan M1013 T+00:00
Scene view
ARM: IDLE
EVS stream
TCP confidence
ROS topic graph
Standby
Rover and Doosan M1013 arm online. ROS2 topic graph ready. EVS camera mounted. Press play to begin simulation.
00:0000:3501:10
ROS master · ready

Data → Algorithm → Hardware

Three layers. One stack. The data moat enables the algorithm. The algorithm targets the hardware. The hardware enables edge deployment at watts-scale power.

Data layer · Phasor Episodes
652 datasets
Spike encoding
7 domains
AER format
v2e · I2E · NIR
Cloudflare R2
Algorithm layer · TCP fusion
EVS + LiDAR + IMU
SNN
Temporal confidence
TCP
Continual learning
Cross-modal
Hardware layer · edge targets
BrainChip Akida
+
Innatera
SynSense Speck
+
Intel Loihi
NIR standard
NeuroBench
Deployment · AV + robotics
ROS2 topics
Arm control
0.8mJ / inference
Edge deploy
Five AI / Bosch
Production

NeuroFusion v0.4 — mission control

Full sensor fusion dashboard — three sensor streams fusing in real time, BEV object tracking with range rings, live performance metrics and system log.

/
NeuroFusion v0.4
LIVE · 1.4MHz spike rate
00:00:00
Sensor streams
Event Camera (EVS)
1.4M ev/s · Δt 0.7µs
LiDAR (3D)
72k pts/s · 128-line
IMU (6-DOF)
1kHz · acc ±0.02g
Fused Output (SNN)
SNN layer 4 · 98.2% conf
Hardware
Target chip
BrainChip Akida
AKD1500 · MetaTF
Mode
Real‑time
Fused perception ·
Performance
2.1ms
fusion latency
848Hz
update rate
94%
energy reduction
8
objects tracked
1.4M
spikes/sec
98.2%
snn confidence
vs GPU baseline:
↓ 94%power
↓ 12×latency
↓ 8×bandwidth
OKfusion pipeline nominal OKSNN layer 4 · 14 active neurons WARNobject at 2.1m — confidence 91% OKAkida AKD1500 · 0.8mJ/inference OKEVS stream nominal · 1.4M ev/s OKLiDAR sync · Δt 0.3ms WARNTCP confidence at 91% — monitoring OKcross-modal fusion nominal · 848Hz OKfusion pipeline nominal OKSNN layer 4 · 14 active neurons WARNobject at 2.1m — confidence 91% OKAkida AKD1500 · 0.8mJ/inference OKEVS stream nominal · 1.4M ev/s OKLiDAR sync · Δt 0.3ms WARNTCP confidence at 91% — monitoring OKcross-modal fusion nominal · 848Hz
↳ 2.1mW
94%

Energy reduction

SNN inference on BrainChip Akida vs GPU equivalent. NICE 2026 & AAAI 2026 peer-reviewed.

2.1ms

Fusion latency

End-to-end cross-modal fusion: EVS + LiDAR + IMU. 847Hz continuous update rate on Akida hardware.

652

Episodes datasets

Event-based datasets across 7 domains — the world’s largest neuromorphic training library.

TCP

Core innovation

Temporal Confidence Propagation — maintains object tracking under occlusion. No standard SNN framework does this.