Sign In

Distributed Multi-Robot Object SLAM using Variational Inference and Consensus Averaging

Core Concepts
This paper presents a distributed approach for multi-robot simultaneous localization and mapping (SLAM) that formulates the problem as a variational inference problem over a communication graph. The method imposes a consensus constraint on the object landmarks maintained by different robots to ensure agreement on a common map, and uses a distributed mirror descent algorithm with a regularization term to enforce this consensus.
The paper addresses the problem of multi-robot SLAM, where a team of collaborating robots aim to jointly estimate their trajectories and build a common map of the environment. The authors formulate this as a variational inference problem over a communication graph, with each robot maintaining a distribution over its own state (poses) and the shared environment (landmarks). The key aspects of the approach are: Formulation as a variational inference problem: The authors cast the multi-robot SLAM problem as minimizing the Kullback-Leibler (KL) divergence between a variational density and the true Bayesian posterior. This allows them to leverage tools from optimization and probabilistic inference. Distributed mirror descent algorithm: The authors develop a distributed mirror descent algorithm to solve the variational inference problem. This involves each robot updating its density based on its own measurements, while also regularizing it to be similar to the densities of its neighbors in the communication graph. Consensus on common landmarks: The authors introduce a consensus constraint on the estimates of common landmarks across robots. This is achieved by having the robots average the marginal densities of the shared landmarks during the mirror descent updates. Distributed multi-state constraint Kalman filter (MSCKF): By using Gaussian distributions in the mirror descent algorithm, the authors derive a distributed version of the MSCKF algorithm, which combines visual odometry, feature observations, and object detections to jointly estimate robot poses and object maps. The paper demonstrates the effectiveness of the proposed approach on both real-world (KITTI dataset) and simulated multi-robot scenarios. The results show that the distributed MSCKF with consensus averaging improves the overall accuracy of trajectories and object maps compared to individual SLAM, while also achieving better scaling to large robot teams compared to centralized approaches.
The paper reports the following key metrics: Trajectory RMSE (meters) on KITTI sequences: Sequence 00: 9.45 (with consensus), 10.81 (without consensus) Sequence 05: 7.33 (with consensus), 9.26 (without consensus) Sequence 06: 4.87 (with consensus), 4.90 (without consensus) Sequence 08: 11.93 (with consensus), 16.18 (without consensus) Object estimation error (meters) on KITTI sequences: Sequence 00: 7.73 (with consensus), 8.36 (without consensus) Sequence 05: 6.87 (with consensus), 7.82 (without consensus) Sequence 06: 4.59 (with consensus), 4.66 (without consensus) Sequence 08: 11.68 (with consensus), 14.08 (without consensus) Object position differences (meters) across robots on KITTI sequences: Sequence 00: 6.56 (with consensus), 9.46 (without consensus) Sequence 05: 5.61 (with consensus), 11.06 (without consensus) Sequence 06: 4.71 (with consensus), 5.12 (without consensus) Sequence 08: 9.12 (with consensus), 19.55 (without consensus) Consensus averaging time per robot per timestep (seconds) in simulation: 3 robots: 0.021 5 robots: 0.022 10 robots: 0.026 15 robots: 0.028
"Centralized processing of robot observations is undesirable because it creates a single point of failure and requires pre-existing infrastructure and significant multi-hop communication throughput." "We impose a consensus constraint on the objects maintained by different nodes to ensure agreement on a common map." "Using Gaussian distributions in the algorithm, we derive a distributed multi-state constraint Kalman filter (MSCKF) for multi-robot object SLAM."

Key Insights Distilled From

by Hanwen Cao,S... at 04-30-2024
Multi-Robot Object SLAM using Distributed Variational Inference

Deeper Inquiries

How could the proposed approach be extended to handle dynamic environments with moving objects

To extend the proposed approach to handle dynamic environments with moving objects, several modifications and enhancements can be implemented. One approach could involve incorporating predictive models for object motion based on historical data and sensor observations. By predicting the future positions of moving objects, the robots can update their object maps and trajectories accordingly. Additionally, integrating techniques from dynamic object tracking and prediction algorithms, such as Kalman filters or particle filters, can help improve the accuracy of object localization in dynamic environments. Furthermore, implementing collaborative tracking and prediction mechanisms among the robots can enhance the overall performance in tracking moving objects across the environment.

What are the theoretical guarantees on the convergence and optimality of the distributed mirror descent algorithm used in this work

The distributed mirror descent algorithm used in this work offers theoretical guarantees on convergence and optimality. Convergence guarantees can be established by showing that the algorithm converges to a stationary point of the objective function. The mirror descent algorithm is known to converge to a solution that is within a certain suboptimality bound of the optimal solution. The convergence rate can be analyzed based on the step size parameter and the properties of the objective function. Additionally, the optimality of the algorithm can be assessed by comparing the achieved solution with the global optimum under certain conditions, such as convexity and smoothness of the objective function.

Can the consensus averaging step be further optimized to reduce communication overhead, for example, by using gossip-based protocols or sparsification techniques

The consensus averaging step can be further optimized to reduce communication overhead by leveraging gossip-based protocols or sparsification techniques. Gossip-based protocols can be utilized to enable efficient information exchange among the robots while minimizing the amount of communication required. By implementing randomized gossip algorithms, the robots can share information with a subset of neighbors, reducing the overall communication load. Sparsification techniques, such as compressive sensing or graph sparsification, can be applied to reduce the amount of information exchanged during the consensus averaging step. By selectively transmitting essential information and discarding redundant data, the communication overhead can be significantly reduced while maintaining the consensus among the robots.