- GARLILEO // ARXIV 2511
- 22,841 REAL SAMPLES
- <2% VERTICAL DRIFT
GNOMON
KNOW WHERE YOU STAND
Legged robots can walk, climb, and navigate — but they can't tell you where they walked. On stairs, position drifts 8–15% vertically per step sequence. GNOMON fuses radar, IMU, and leg kinematics to cut vertical error to under 2%. Validated on 22,841 real sensor samples.
MODULE STATUS: PROVEN MVPDataset
<2%
- DIVISION
- ANIMA
- WAVE
- W3
- DOMAIN
- FOUNDATION
- MODULE 05 // PROVEN MVP
- RADAR-LEG-INERTIAL ODOMETRY
LEGGED ROBOTS DON'T KNOW WHERE THEY ARE ON STAIRS
Quadrupeds and humanoids are shipping — Boston Dynamics, Unitree, Figure, Agility. They can walk. They can climb. But the moment they hit stairs, slopes, or rough terrain, their position estimate drifts by 8–15% vertically per step sequence.
Wheel odometry doesn't exist on legs. GPS fails indoors. Cameras break in dust, rain, and darkness. IMU alone accumulates error. There is no reliable "where am I" for legged robots in real-world terrain. Robots can walk. They just can't tell you where they walked.
WHAT GNOMON DELIVERS
GNOMON fuses three sensor streams that legged robots already carry into a continuous-time optimization that keeps vertical error below 2%.
PIPELINE
- 01RADAR DOPPLERGround-relative velocity that anchors the estimate even when legs slip
- 02IMU (6-DOF)Angular rate and acceleration — high-frequency motion sensing
- 03LEG KINEMATICSJoint encoder readings from the robot's own body
CAPABILITIES
- Continuous-time spline optimization via Ceres Solver (C++ core)
- Gravity alignment decouples vertical behavior from heading drift
- Sub-2% vertical accuracy where nothing else works
- End-to-end latency under 50ms, fusion rate up to 100 Hz
- 25+ REST API endpoints for replay, benchmarking, and operations
- Platform configs: quadruped (Spot, Go2, B2) and humanoid (custom DH chains)
WHY THIS IS HARD
Every sensor alone is unreliable on stairs:
- 01IMU alone: drifts over time, accumulates bias — no absolute reference
- 02Kinematics alone: 8–15% vertical error on stairs (legs slip, contact timing varies)
- 03Cameras alone: fail in repetitive environments, dust, darkness, rain
- 04GPS: doesn't exist indoors or underground — not an option
- 05The breakthrough: radar Doppler fusion — a modality most robotics companies ignore
Radar provides ground-relative velocity that doesn't care about lighting, dust, or repetitive textures. Combined with IMU and leg kinematics in a continuous-time spline framework, you get sub-2% vertical accuracy where nothing else works.
REAL SENSOR VALIDATION
Validated replay of official GaRLILEO dataset:
| METRIC | VALUE |
|---|---|
| Dataset | Official GaRLILEO Upstair ROS 2 bag |
| Pose Samples Recovered | 22,841 |
| Replay Duration | 228.4 seconds |
| Recovered Path Length | 189.97 meters |
| Vertical Drift Target | <2% (vs 8–15% kinematics-only) |
| End-to-End Latency | <50ms |
| Fusion Rate | Up to 100 Hz |
| Unit Tests Passing | 88 |
| Lint/Type Errors | 0 |
WHAT'S BUILT TODAY
| COMPONENT | STATUS | NOTES |
|---|---|---|
| Core models | COMPLETE | C++ fusion engine (Ceres + Eigen3) — production-ready |
| C++ Fusion Engine | COMPLETE | Continuous-time spline optimization, Ceres Solver |
| Python API Wrapper | COMPLETE | FastAPI, 25+ endpoints |
| Replay Control Plane | COMPLETE | Jobs, logs, cancel, timeout management |
| Trajectory Visualization | COMPLETE | SVG export, Pangolin rendering |
| Operator Dashboard | COMPLETE | /operator/overview — live operations |
| Benchmark Engine | COMPLETE | Cross-run comparison and reporting |
| Docker Build | COMPLETE | Multi-stage, ROS 2 Humble base |
| Dataset Validation | COMPLETE | Topic + structure checking |
| Test Suite | COMPLETE | 88 tests, 0 lint/type errors |
| API layer | IN PROGRESS | Public API — pending dataset infrastructure |
| ROS2 Live Topics | PLANNED | Live sensor streaming instead of bag replay |
| TI mmWave Radar | PLANNED | Hardware integration for live radar fusion |
SUPPORTED PLATFORMS
- Quadruped (Spot, Go2, B2)Configuration ready — standard leg kinematics
- Humanoid (custom DH chains)Configuration ready — adaptable kinematics model
- Custom KinematicsTOML adapter — define your own robot model
WHERE GNOMON DEPLOYS
- APP_01
LEGGED ROBOT MANUFACTURERS
Unitree, Boston Dynamics, Agility, Figure — every legged robot needs accurate odometry.
- APP_02
CONSTRUCTION SITE AUTONOMY
Stairs, scaffolding, slopes — GPS-denied environments where accurate positioning is critical.
- APP_03
DISASTER RESPONSE
Unstructured terrain with dust, debris, and darkness. Radar doesn't care.
- APP_04
UNDERGROUND & MINING
GPS-denied, camera-hostile environments. Radar+IMU+kinematics is the only option.
- APP_05
MULTI-FLOOR DELIVERY
Last-mile delivery in buildings with stairs and elevators.
- APP_06
RESEARCH INSTITUTIONS
GaRLILEO baseline for radar-inertial-leg odometry benchmarking.
UNDER THE HOOD
FOUNDATION: GARLILEO (ARXIV 2025)
- Gravity-aligned Radar-Leg-Inertial Enhanced Odometry
- Continuous-time spline optimization (C++ / Ceres Solver 2.2.0 / Eigen3)
- Radar Doppler: ground-relative velocity immune to visual conditions
- Gravity alignment: decouple vertical from heading drift
- Validated on official GaRLILEO Upstair dataset (22,841 samples)
FUSION ARCHITECTURE
- Radar Doppler → Continuous-time spline → Fused odometry + gravity
- IMU (6-DOF) → Angular rate + acceleration integration
- Leg kinematics → Joint encoder readings → Forward kinematics
- Foot contacts → Contact timing → Stance phase detection
- Output: REST API / gRPC → Fleet integration
API ENDPOINTS
- GET /odom
- Current pose, velocity, covariance
- POST /replay/jobs
- Start a replay job from ROS 2 bag
- GET /telemetry
- Sensor health (IMU/radar/kinematics Hz)
- GET /benchmarks/report
- Cross-run benchmark comparison
RESEARCH PAPER
- [01]GaRLILEO: Gravity-aligned Radar-Leg-Inertial Enhanced Odometry — arXiv 2511.13216, MIT, Nov 2025