Skip to content
RFL_GLOBAL
中文
BACK TO PLATFORM
ROBOT FLOW LABSPOINT CLOUD REGISTRATION

BIFROST

CUDA REGISTRATION ENGINE

60+ FPS. 98.4% KITTI recall. The registration engine we built because Open3D wasn’t fast enough.

KITTI THROUGHPUT

67.8 FPS

  • 98.4% RECALL
  • 13× OPEN3D
  • 8+ SLAM MODULES
BIFROST // POINT CLOUD REGISTRATION
01THE BOTTLENECK

THE PROBLEM

Point cloud registration is the geometric glue for all SLAM. Every time a LiDAR sweeps or a depth camera fires, the system must align the new scan to the previous one. This is registration.

Open3D ICP runs at ~5 FPS. Modern LiDAR operates at 10–30 Hz. That’s catastrophically slow — your registration pipeline can’t keep up with your sensor.

RANSAC is stochastic and non-deterministic. Run it twice, get two different answers. Unacceptable for safety-critical robotics.

RobotFlow Labs ships 8+ SLAM modules. We couldn’t build a real SLAM suite on a 5 FPS bottleneck.

02THE CORE ALGORITHM

TURBOCLIQUE

Core insight: instead of random sampling (RANSAC), build a consistency graph from putative correspondences, find maximal 3-cliques on GPU, and use them to filter before pose solve.

  1. 01

    CORRESPONDENCE GRAPH

    Build a compatibility graph from feature matches. Edges encode pairwise geometric consistency.

  2. 02

    3-CLIQUE DETECTION

    Find maximal 3-cliques on GPU. Each clique is a triplet of mutually consistent correspondences — a strong pose hypothesis.

  3. 03

    GPU FILTERING

    Filter thousands of cliques in parallel on CUDA. Deterministic. No random sampling. Same input always produces same output.

  4. 04

    POSE SOLVE

    Weighted SVD from filtered correspondences. Sub-degree rotation accuracy. Centimeter-level translation.

03THE CORE ALGORITHM

ARCHITECTURE

C++ CUDA core (.so binaries, proprietary) + Python layer (open source). Same distribution model as cuDNN / TensorRT.

  • PYTHON API

    OPEN

    Open source — bifrost.register()

  • C++ CUDA CORE

    PROPRIETARY

    Proprietary — compiled .so binaries

  • GPU RUNTIME

    HARDWARE

    NVIDIA CUDA — Ampere / Hopper

04PERFORMANCE

KITTI ODOMETRY

OUTDOOR LIDAR

98.4%
REGISTRATION RECALL
0.40°
ROTATION ERROR
8.12 cm
TRANSLATION ERROR
67.8 FPS
THROUGHPUT

13× faster than Open3D ICP

05PERFORMANCE

3DMATCH

INDOOR RGB-D

93.6%
RECALL (FCGF)
2.04°
ROTATION ERROR
6.42 cm
TRANSLATION ERROR
64.6 FPS
THROUGHPUT

With FCGF feature backbone

06PERFORMANCE

HEAD TO HEAD

  • BIFROST60–68 FPSCUDA + TurboClique
  • OPEN3D ICP RANSAC~5 FPSCPU

12–13× SPEEDUP

07SLAM SUITE

POWERS EVERY SLAM MODULE

  • PRISM

    Multi-sensor Gaussian SLAM

  • GS3LAM

    3D Gaussian Splatting SLAM

  • CokO

    Multi-agent collaborative SLAM

  • MipSLAM

    Mip-NeRF based SLAM

  • INTACT

    LiDAR integrity monitoring

  • Ghost-FWL

    Phantom return rejection

  • ProjFusion

    Camera-LiDAR fusion

  • OccAny

    3D occupancy prediction

None of these ship Open3D ICP. All of them ship BIFROST.

08INTEGRATION

DEPLOYMENT

  • C++ CUDA CORE

    Compiled .so binaries for Linux x86_64 + ARM64

  • PYTHON API

    bifrost.register(source, target, features)

  • CLI

    bifrost-cli register --source a.pcd --target b.pcd

  • REST API

    FastAPI server for network-accessible registration

  • DOCKER

    CUDA-ready container with all dependencies

  • PRESETS

    10+ dataset-specific configs: KITTI, 3DMatch, nuScenes…

STATUS

SHIPPING

  • C++ CUDA core shipping. Powers all 8 SLAM modules in production.
  • Proprietary IP distributed as compiled .so — same commercial model as cuDNN.
  • Python API open source. Core engine closed source.

PAPERTurboReg — ICCV 2025

REQUEST ACCESSBACK TO PLATFORM