Skip to content
RFL_GLOBAL
中文
  • MetricAnything // 2601.22054
  • GPU-train, MLX inference
  • WAVE 6

ODIN

ALL-SEEING DEPTH

ODIN delivers absolute metric depth from a single RGB image by pre-training on 20 million heterogeneous 3D data sources — LiDAR, stereo, RGB-D, and Structure-from-Motion. Unlike relative depth estimators that only rank pixels by distance, ODIN outputs real-world measurements in meters, enabling robots to judge distances with centimeter-level accuracy without any manual calibration. The system uses sparse metric prompts to adapt across sensor types, making it the foundational depth backbone for the entire ANIMA perception stack. Every downstream module — from SLAM to manipulation planning — depends on ODIN's accurate distance measurements.

MODULE STATUS: DEVELOPMENT

Training Sources

20M

DIVISION
ANIMA
WAVE
W6
DOMAIN
DEPTH SYSTEMS
WAVE 6 // ANIMA SUITE
FOUNDATION — METRIC DEPTH
ODIN // W6 // 056/079
01THE CHALLENGE

DEPTH WITHOUT GROUND TRUTH

Robots need exact distances in meters, not relative ordering. Traditional depth sensors are expensive and fail in many conditions.

Every downstream module depends on accurate distance measurements.

02THE SOLUTION

WHAT ODIN DELIVERS

ODIN delivers absolute metric depth from a single RGB image by pre-training on 20 million heterogeneous 3D data sources. Unlike relative depth estimators, ODIN outputs real-world measurements in meters.

CAPABILITIES

  • Pre-trained on 20M heterogeneous 3D sources
  • Outputs absolute metric depth in meters
  • Sparse metric prompts for cross-sensor adaptation
  • Foundation backbone for ANIMA perception stack
03ENGINEERING

WHY THIS IS HARD

Building ODIN requires solving multiple coupled problems:

  1. 01Unifying heterogeneous depth sources with different scales and noise
  2. 02Learning absolute scale without per-scene calibration
  3. 03Maintaining accuracy across environments
  4. 04Real-time inference for robotics

ODIN solves these through careful architecture design and rigorous validation.

04BENCHMARKS

PROOF, NOT PROMISES

Key metrics:

PROOF, NOT PROMISES
METRICVALUE
Training Sources20M+
OutputMetric (meters)
CalibrationNone
BackendGPU + MLX
05BUILD STATUS

WHAT'S BUILT TODAY

3/6 COMPONENTS COMPLETE
WHAT'S BUILT TODAY
COMPONENTSTATUSNOTES
Pre-training PipelineCOMPLETE20M multi-modal pre-training
Multi-Source FusionCOMPLETELiDAR+stereo+RGB-D unified
Metric PromptingIN PROGRESSCross-sensor adaptation
Edge InferenceIN PROGRESSMLX optimization
Core modelsCOMPLETEFoundation depth validated
API layerIN PROGRESSREST + streaming
06APPLICATIONS

WHERE ODIN DEPLOYS

  • APP_01

    AUTONOMOUS NAVIGATION

    Metric depth for obstacle avoidance without LiDAR.

  • APP_02

    SLAM BACKBONE

    Foundation depth for THOR, BALDUR, HEIMDALL.

  • APP_03

    MANIPULATION

    Distance measurements for grasp planning.

07TECHNOLOGY

UNDER THE HOOD

FOUNDATION: METRICANYTHING

  • Pre-trained on 20M heterogeneous 3D sources
  • Outputs absolute metric depth in meters
  • Sparse metric prompts for cross-sensor adaptation

KEY INNOVATION

ODIN delivers absolute metric depth from a single RGB image by pre-training on 20 million heterogeneous 3D data sources. Unlike relative depth estimators, ODIN outputs real-world measurements in meters.

DEPLOYMENT

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

COMPUTE

PRIMARY
GPU-train, MLX inference
EDGE
Optimized inference
API
REST + streaming
08PAPERS

PAPERS

  1. [01]MetricAnything (2601.22054)