- CALIB-PROJFUSION // W7
- Python
- ANIMA
CALIB-PROJFUSION
Align Your Sensors Once. Trust Them Forever.
Camera-LiDAR sensor fusion is only as good as the extrinsic calibration holding it together. In field conditions — vehicle vibration, rough terrain landings, temperature swings — that calibration drifts. Recalibration in the field is slow, requires controlled environments, and often gets skipped. Misaligned sensors produce fused outputs that are confidently wrong. CALIB-PROJFUSION implements ProjFusion (RA-L 2026, arXiv:2603.29414), a native-domain cross-attention architecture that reasons jointly over a DINOv2-ViT-S/14 image encoder and a PointGPT-tiny point cloud encoder.
MODULE STATUS: PRODUCTIONRobustness
PROD
- DIVISION
- GENERIC
- WAVE
- W7
- DOMAIN
- GENERAL
- WAVE 7 // ANIMA SUITE
- FOUNDATION — CALIB-PROJFUSION
THE PROBLEM WE SOLVE
Camera-LiDAR sensor fusion is only as good as the extrinsic calibration holding it together. In field conditions — vehicle vibration, rough terrain landings, temperature swings — that calibration drifts. Recalibration in the field is slow, requires controlled environments, and often gets skipped. Misaligned sensors produce fused outputs that are confidently wrong.
Autonomous ground vehicles, UAVs, and robotic systems can self-recalibrate in the field after sensor disturbance, maintaining fusion accuracy without a depot visit or a controlled calibration target.
WHAT CALIB-PROJFUSION DELIVERS
CALIB-PROJFUSION implements ProjFusion (RA-L 2026, arXiv:2603.29414), a native-domain cross-attention architecture that reasons jointly over a DINOv2-ViT-S/14 image encoder and a PointGPT-tiny point cloud encoder. Crucially, it handles large initial perturbations — the kind that result from a hard landing or a field swap — without iterative refinement.
CAPABILITIES
- CALIB-PROJFUSION implements ProjFusion (RA-L 2026, arXiv:2603.29414), a native-domain cross-attention architecture that reasons jointly over a DINOv2-ViT-S/14 image encoder and a PointGPT-tiny point cloud encoder
- Crucially, it handles large initial perturbations — the kind that result from a hard landing or a field swap — without iterative refinement
- The 6D pose is output as an SE(3) Lie algebra vector, directly usable by any robotics pose chain
- Only 3.6M of the 25.3M parameters are trained.
WHY THIS IS HARD
Building CALIB-PROJFUSION requires solving multiple coupled problems:
- 01Camera-LiDAR sensor fusion is only as good as the extrinsic calibration holding it together
- 02In field conditions — vehicle vibration, rough terrain landings, temperature swings — that calibration drifts
- 03CALIB-PROJFUSION implements ProjFusion (RA-L 2026, arXiv:2603.29414), a native-domain cross-attention architecture that reasons jointly over a DINOv2-ViT-S/14 image encoder and a PointGPT-tiny point cloud encoder
- 04Crucially, it handles large initial perturbations — the kind that result from a hard landing or a field swap — without iterative refinement
CALIB-PROJFUSION solves these through careful architecture design and rigorous validation.
PROOF, NOT PROMISES
Key performance metrics:
| METRIC | VALUE |
|---|---|
| Robustness | Handles large initial perturbations — no iterative alignment or calibration target needed |
| Efficiency | 3.6M trainable parameters from a 25.3M total architecture — fast inference, small update cost |
| Input | 3×224×448 image + 8,192 LiDAR points — lightweight sensor requirement |
| Defense Angle | Field-resilient sensor fusion for UGVs, UAVs, and robotic platforms operating in contested terrain |
WHAT'S BUILT TODAY
| COMPONENT | STATUS | NOTES |
|---|---|---|
| Robustness | COMPLETE | Handles large initial perturbations — no iterative alignment or calibration target needed |
| Efficiency | COMPLETE | 3.6M trainable parameters from a 25.3M total architecture — fast inference, small update cost |
| Input | IN PROGRESS | 3×224×448 image + 8,192 LiDAR points — lightweight sensor requirement |
| Defense Angle | IN PROGRESS | Field-resilient sensor fusion for UGVs, UAVs, and robotic platforms operating in contested terrain |
WHERE CALIB-PROJFUSION DEPLOYS
- APP_01
AUTONOMOUS SYSTEMS
Autonomous ground vehicles, UAVs, and robotic systems can self-recalibrate in the field after sensor disturbance, maintaining fusion accuracy without a depot visit or a controlled calibration target.
- APP_02
RESEARCH LABS
CALIB-PROJFUSION implements ProjFusion (RA-L 2026, arXiv:2603.29414), a native-domain cross-attention architecture that reasons jointly over a DINOv2-ViT-S/14 image encoder and a PointGPT-tiny point cloud encoder.
- APP_03
EDGE COMPUTING
25.3M-parameter camera-LiDAR calibration engine that self-corrects after hard landings, vibration, or thermal drift — no checkerboard required.