Source-linked AI summary
Data-Efficient Decentralized Visual SLAM
Titus Cieslewski, Siddharth Choudhary, Davide Scaramuzza
TL;DR
Multi-robot visual SLAM needs decentralized operation without relying on a central entity, while prior work provided key data-efficient components separately. This paper integrates those components into a complete decentralized visual SLAM system, characterizes its behavior on public data, and shows that bandwidth can be reduced without significantly affecting reported accuracy.
Problem
Visual multi-robot SLAM must support teams operating without absolute positioning while avoiding dependence on a central entity that must remain reachable, reliable, and scalable.
Method
The paper integrates decentralized place recognition, visual feature association, relative measurements, and optimization into a complete decentralized visual SLAM system.
Results
The system was evaluated on publicly available data, and increasing τmdg significantly throttled bandwidth without a significant effect on ATE, which remained around 4m.
Takeaways & Limitations
The results support further work on more data-efficient data association and more robust decentralized optimization.
Abstract
from arXiv · showhide
Decentralized visual simultaneous localization and mapping (SLAM) is a powerful tool for multi-robot applications in environments where absolute positioning systems are not available. Being visual, it relies on cameras, cheap, lightweight and versatile sensors, and being decentralized, it does not rely on communication to a central ground station. In this work, we integrate state-of-the-art decentralized SLAM components into a new, complete decentralized visual SLAM system. To allow for data association and co-optimization, existing decentralized visual SLAM systems regularly exchange the full map data between all robots, incurring large data transfers at a complexity that scales quadratically with the robot count. In contrast, our method performs efficient data association in two stages: in the first stage a compact full-image descriptor is deterministically sent to only one robot. In the second stage, which is only executed if the first stage succeeded, the data required for relative pose estimation is sent, again to only one robot. Thus, data association scales linearly with the robot count and uses highly compact place representations. For optimization, a state-of-the-art decentralized pose-graph optimization method is used. It exchanges a minimum amount of data which is linear with trajectory overlap. We characterize the resulting system and identify bottlenecks in its components. The system is evaluated on publicly available data and we provide open access to the code.
I. INTRODUCTION
Multi-robot visual SLAM can accelerate exploration and mapping without relying on a central entity, but decentralized systems must manage communication, consistency, and data-association challenges. This work integrates decentralized place recognition, optimization, and visual feature association into a complete system evaluated on public data with open-source code.
- Multi-robot systems can accelerate exploration and mapping or support heterogeneous teams with specialized robots.
- Centralized visual SLAM requires a reachable, reliable, and sufficiently scalable central entity for computation and bandwidth.
- The paper integrates state-of-the-art decentralized place recognition and optimization with data-efficient visual feature association into a complete system.The system is evaluated on publicly available data, and its code is provided open-source.
- Distributed estimation research emphasizes communication constraints, heterogeneity, consistency, and robust data association across filtering and smoothing approaches.
- Decentralized trajectory estimation addresses centralized communication and computation constraints by using local communication to reach trajectory consensus.
B. Decentralized Place Recognition and Pose Estimation
The system establishes inter-robot constraints through indirect visual observations rather than specialized direct sensors, while combining decentralized place recognition, relative pose estimation, visual odometry, and optimization. Its design supports continuous operation, with stereo visual odometry providing absolute scale.
- Decentralized Place Recognition and Pose Estimation: Direct inter-robot measurements often require specialized hardware and line of sight, limiting the relative poses that can be established.The paper notes that this limitation can constrain higher-level applications using decentralized SLAM.
- Decentralized Place Recognition and Pose Estimation: Indirect relative measurements register observations of the same scene, reusing odometry sensors without additional hardware.They remain limited by communication range, but that range can exceed line of sight in practice.
- Decentralized Place Recognition and Pose Estimation: Each robot runs NetVLAD for compact image representations, visual odometry for a sparse feature map, and RelPose for candidate-match verification and relative pose estimation.
- Decentralized Place Recognition and Pose Estimation: The system updates its map continuously as new images arrive and relative pose measurements are produced.
- Intra-Robot Measurements: Visual odometry must produce a pose graph, associate each pose with an image, and provide absolute scale for the system.The implementation uses stereo ORB-SLAM; monocular odometry requires scale recovery from stereo or inertial measurements.
B. Inter-Robot Measurements
Inter-robot measurements are established in two phases: compact decentralized place recognition first, followed conditionally by relative pose estimation. The first phase selects one robot for the query and avoids unnecessary data exchange.
- Two-phase measurement establishment: Place recognition first tests whether another robot captured the same scene, before requesting relative pose estimation.Relative pose estimation is more data-intensive, so the first phase filters and routes subsequent queries.
- Two-phase measurement establishment: NetVLAD represents each image as a full-image descriptor whose distances tend to be lower for images observing the same scene.The descriptor is a vector vαi ∈ R^DNetVLAD computed by a neural network.
- Communication-efficient routing: The descriptor query is sent only to robot δ, selected by the nearest cluster center, rather than to every robot.Robot δ already stores prior queries assigned to its Voronoi cell and can return a matching image identifier without extra exchange.
- Communication-efficient routing: The routing method requires |Ω| times less data exchange than sending the query to all other robots.Cluster centers are obtained using k-means clustering on the Oxford Robotcar Dataset.
D. Relative Pose Estimation
Relative pose estimation uses compact landmark and visual-word data, reuses visual-odometry products, and applies geometric verification with thresholds to accept consistent matches while limiting communication.
- Relative pose computation: Relative pose estimation reuses visual-odometry products such as landmark 3D positions and imitates ORB-SLAM’s loop-closure method without exchanging surrounding match data.The implementation is designed to avoid the larger data exchange that surrounding observations would require.
- Relative pose computation: The transmitted representation associates keypoints with visual-word identifiers and landmark positions expressed in the source camera frame.Visual words quantize ORB descriptor space and support keypoint matching between images.
- Geometric verification: RANSAC estimates an initial relative pose from matched landmark pairs, and inlier pairs refine the pose through robust 3D registration.A place match is rejected when the inlier count is below τinliers.
- Geometric verification: The robust arctan loss reduces the weight of remaining keypoint-match outliers.The loss uses the parameter τloss.
- Geometric verification: Geometric verification is throttled by τmdg, skipping matches close to already established relative poses.This reduces repeated transmission when an existing accepted match provides a nearby verification context.
E. Decentralized Optimization
Decentralized optimization estimates robot trajectories from relative-pose measurements through a two-stage distributed solver. Its block-structured linear systems are solved with distributed Gauss-Seidel updates that converge to the original system’s solution when the iterations converge.
- Optimization formulation: Each robot trajectory is represented as a sequence of poses xαi=(Rαi,tαi) in SE(3), with relative measurements linking pose pairs.Rotation noise is modeled with a Von-Mises distribution and translation noise with a zero-mean Gaussian.
- Two-stage optimization: The distributed solver first estimates all rotations from a relaxed optimization problem, then recovers full poses and performs a Gauss-Newton iteration.This is the stated two-stage approach for solving the optimization problem in a distributed manner.
- Distributed linear systems: The first two linear systems partition unknowns and equations into subvectors and blocks associated with individual robots.They are expressed in the general form Hy=g, with H and g partitioned according to y’s block structure.
- Distributed linear systems: Distributed Gauss-Seidel updates each robot’s subvector using the latest estimates from robots that have already updated.The iteration begins from an arbitrary initial estimate y(0).
- Distributed linear systems: If the distributed Gauss-Seidel sequence converges to a fixed point, that point satisfies the block equations and solves the original linear system.The convergence statement connects the distributed iteration to the centralized linear-system solution.
F. Making Optimization Work Continuously
Continuous operation requires coordinating visual odometry, place recognition, and episodic decentralized optimization because intermediate optimization maps are not always consistent. Optimistic concurrency control merges optimization results safely with the reference state.
- Consistency during continuous operation: Intermediate and first-stage optimization maps are not consistent states, so DOpt’s map cannot be treated as the best estimate at every time.The optimization also is not designed to incorporate new measurements between iterations.
- Consistency during continuous operation: Concurrent modification by visual odometry and decentralized optimization could create inconsistencies unless their updates are controlled.The method treats VO and place recognition as one process and decentralized optimization as another.
- Optimistic concurrency control: Optimistic concurrency control gives concurrent processes separate copies and merges a completed process’s copy back into the reference state.In this system, decentralized optimization runs episodically after the robots agree on a reference time.
- Optimistic concurrency control: The correction applies to remaining poses xαj recorded after the agreed reference time.The frame immediately before that time is denoted (α,e).
IV. EXPERIMENTS
The experiments measure data use by individual components and the overall system alongside state-estimate accuracy over time, assessing applicability, scalability, and optimization priorities.
- The experiments measure component-level and overall data usage together with state-estimate accuracy as time progresses.
A. Data Exchange Evaluation
The data-exchange evaluation records transfers for decentralized place recognition, relative pose estimation, and decentralized optimization, using explicit byte-size accounting for exchanged elements.
- Data transfer is recorded separately for decentralized visual place recognition, relative pose estimation, and decentralized optimization.
- The exchange-size accounting includes poses, NetVLAD vectors, robot indices, frame indices, visual word indices, and landmarks.The stated sizes are 6 · 8 bytes for poses, 8·DNetVLAD bytes for NetVLAD vectors, 1 byte for robot indices, 4 bytes for frame indices, 2 bytes for visual word indices, and 3 · 4 bytes for landmarks.
B. Accuracy Evaluation
The accuracy evaluation tracks how state-estimate accuracy evolves over time using average trajectory error, reported separately for connected components.
- Accuracy is measured with average trajectory error (ATE), whose ground-truth alignment is meaningful only for connected components.
- ATE is reported separately for different connected components to characterize accuracy over time.
C. Parameter Studies
The parameter studies characterize communication and accuracy trade-offs across decentralized optimization, place recognition, and relative pose estimation. They show that bandwidth can be reduced substantially while maintaining accuracy, but traffic remains uneven and some parameter choices increase relative-pose communication.
- Communication breakdown: RelPose and DOpt require the most data transmission, while DOpt traffic can be significantly reduced and RelPose traffic remains high.The system’s component traffic is evaluated over time and between robot pairs.
- Communication breakdown: Higher trajectory overlap increases DOpt and RelPose traffic, while DVPR traffic is uneven because NetVLAD cluster centers are assigned to specific robots.The unevenness is visible in pairwise transmission and is attributed to clustering; robots 2, 6, and 7 receive centers representing most KITTI 00 descriptors.
- Accuracy: ATE stabilizes around 4 meters after connected components merge, remaining below 1% of the overall trajectory length.Early merges can increase ATE, whereas later loop closures reduce it again before stabilization.
- Relative-pose parameter: Increasing τmdg significantly reduces bandwidth requirements without negatively affecting accuracy.τmdg controls the distance between an established relative pose and the next frame sent for geometrical verification.
- Place-recognition parameter: The ATE remains around 4m as NetVLAD dimensionality varies, while too-low dimensionality lowers precision and increases failed relative-pose estimates.An efficient trade-off is reached with DNetVLAD > 100; beyond a certain point, larger dimensions do not significantly improve precision.
- Additional evaluation: The MIT Stata Center evaluation produced a final estimate that was not visually similar to the groundtruth estimate.The supplied evaluation passage identifies this as a limitation of the result.
VI. CONCLUSION
The paper presents an integrated decentralized visual SLAM algorithm, characterizes its data transmission, and identifies directions for more efficient association and more robust optimization. Evaluation includes ten MIT Stata Center sub-trajectories before and after final loop closure.
- The authors present a new integrated decentralized visual SLAM algorithm based on state-of-the-art components.
- The system is characterized using publicly available data while exploring how data transmission can be reduced to a minimum.
- Future developments should pursue more data-efficient data association and more robust decentralized optimization.
- Ten MIT Stata Center sub-trajectories are shown before and after the final loop closure, with colors identifying individual robot trajectories.