Skip to content
RFL_GLOBAL
中文
  • AsyncMDE // 2603.10438
  • MLX-OK
  • WAVE 6

GRID

THE ACCELERATOR

GRID achieves 237 FPS depth estimation on RTX 4090 and 161 FPS on Jetson AGX Orin with only 3.83 million parameters — a 25x parameter reduction over standard models. It runs an expensive foundation model like ODIN only on keyframes (10-15% of frames) and propagates depth at full framerate via a lightweight asynchronous spatial memory network. Foundation depth models are too slow for real-time robotics — GRID amortizes the computational cost over time, enabling edge-deployable depth perception without sacrificing accuracy. Critical for drones, mobile robots, and any platform with strict power and compute budgets.

MODULE STATUS: DEVELOPMENT

FPS (4090)

237FPS

DIVISION
ANIMA
WAVE
W6
DOMAIN
DEPTH SYSTEMS
WAVE 6 // ANIMA SUITE
FOUNDATION — REAL-TIME EDGE DEPTH
GRID // W6 // 030/079
01THE CHALLENGE

FOUNDATION MODELS ARE TOO SLOW

Foundation depth models produce great results but run at 5-10 FPS. Real-time robotics needs 30+ FPS.

Edge hardware has strict power and compute budgets that large models exceed.

02THE SOLUTION

WHAT GRID DELIVERS

GRID achieves 237 FPS depth on RTX 4090 by running expensive models on keyframes only and propagating depth via async spatial memory.

CAPABILITIES

  • 237 FPS RTX 4090, 161 FPS Jetson AGX Orin
  • Only 3.83M parameters — 25x reduction
  • Keyframe-based expensive model amortization
  • Async spatial memory for full-rate propagation
03ENGINEERING

WHY THIS IS HARD

Building GRID requires solving multiple coupled problems:

  1. 01Maintaining accuracy while amortizing expensive computation
  2. 02Async spatial memory coherence across frames
  3. 03Keyframe selection strategy
  4. 04Edge deployment with tight power budgets

GRID solves these through careful architecture design and rigorous validation.

04BENCHMARKS

PROOF, NOT PROMISES

Key metrics:

PROOF, NOT PROMISES
METRICVALUE
FPS (4090)237
FPS (Jetson)161
Parameters3.83M
Reduction25x
05BUILD STATUS

WHAT'S BUILT TODAY

3/6 COMPONENTS COMPLETE
WHAT'S BUILT TODAY
COMPONENTSTATUSNOTES
Keyframe ExtractionCOMPLETESmart keyframe selection
Async MemoryCOMPLETESpatial memory network
Full-Rate PropIN PROGRESSFull framerate depth
Edge OptimizationIN PROGRESSJetson/MLX tuning
Core modelsCOMPLETEEdge depth validated
API layerIN PROGRESSStreaming depth service
06APPLICATIONS

WHERE GRID DEPLOYS

  • APP_01

    REAL-TIME ROBOTICS

    Full-speed depth for reactive navigation.

  • APP_02

    EDGE DEPLOYMENT

    Deploy on Jetson with minimal power.

  • APP_03

    HIGH-SPEED SYSTEMS

    Depth at 200+ FPS for fast-moving platforms.

07TECHNOLOGY

UNDER THE HOOD

FOUNDATION: ASYNCMDE

  • 237 FPS RTX 4090, 161 FPS Jetson AGX Orin
  • Only 3.83M parameters — 25x reduction
  • Keyframe-based expensive model amortization

KEY INNOVATION

GRID achieves 237 FPS depth on RTX 4090 by running expensive models on keyframes only and propagating depth via async spatial memory.

DEPLOYMENT

  • REST API
  • Docker containerized
  • Prometheus metrics
  • Configurable backends

COMPUTE

PRIMARY
MLX-OK
EDGE
Optimized inference
API
REST + streaming
08PAPERS

PAPERS

  1. [01]AsyncMDE (2603.10438)