Cooperative Localization

The core data exchange in cooperative localization involves robots sharing relative measurements of each other. This typically includes:

• Range: Distance between robots, measured via UWB, RF signal strength, or time-of-flight.
• Bearing: Relative angle, measured with cameras or directional antennas.
• Relative Pose: Full relative position and orientation, estimated via visual odometry or LiDAR scan matching. These measurements provide constraints that tightly couple the state estimates of the team, reducing individual uncertainty.

Unlike a centralized filter, cooperative localization uses distributed algorithms to fuse shared data. Each robot maintains its own local estimate of the team's state. Common frameworks include:

• Distributed Kalman Filters (DKF): Variants like the Consensus Kalman Filter where robots iteratively average estimates with neighbors.
• Particle Filters: Used in Monte Carlo localization methods, where particles represent possible poses.
• Factor Graph Optimization: A Graph-SLAM approach where relative measurements create constraints (factors) between robot pose nodes, optimized collectively. This distribution enhances scalability and robustness to single-point failures.

System performance is dictated by the communication graph. Key constraints include:

• Intermittent Links: Real-world wireless networks have dropouts, requiring algorithms tolerant to delayed or lost data.
• Limited Bandwidth: Algorithms must minimize data exchange, often sharing only covariance-intersected summaries or innovations.
• Graph Connectivity: The team must maintain a sufficiently connected network for information to propagate. A spanning tree is often the minimum requirement for global observability. Algorithms are designed for peer-to-peer or multi-hop communication, not just broadcast to a central hub.

For the team's global pose to be estimable, the system must be observable. A key insight: absolute positioning information must enter the system. This is achieved through:

• Anchors: A subset of robots with access to GPS, fixed beacons, or known starting locations.
• Intermittent Absolute Updates: Even occasional GPS fixes for one robot can propagate through the network, correcting long-term drift for all. Without anchors, the team can only estimate relative positions and is susceptible to collective drift. The required number of anchors depends on the measurement types and graph connectivity.

A primary advantage is graceful degradation. The system is designed to tolerate failures:

• Robot Dropout: If a robot fails, the remaining team continues localizing, though with potentially increased uncertainty.
• Sensor Degradation: A robot with a faulty IMU can rely on relative measurements from peers to stabilize its estimate.
• Spoofing/Malicious Data: Advanced implementations incorporate Byzantine fault tolerance to identify and reject incorrect data from compromised agents. This makes it suitable for missions in GPS-denied, hazardous, or adversarial environments like search-and-rescue or military reconnaissance.

Implementations navigate key engineering trade-offs:

• Accuracy vs. Communication: More frequent data sharing improves accuracy but consumes bandwidth and power.
• Centralized vs. Distributed: Centralized filters (e.g., a single EKF for the whole team) are optimal but create a bottleneck. Distributed methods sacrifice some optimality for robustness.
• Consistency: Poorly handled correlations between estimates can lead to overconfidence (incorrectly small covariance). Algorithms like Covariance Intersection are used to maintain consistent, conservative estimates.
• Scalability: Computation per robot should be sub-linear in team size. Consensus-based methods often scale well, while dense factor graph optimization can become costly.

The foundational computational problem of constructing a map of an unknown environment while simultaneously tracking the agent's location within it. Cooperative localization extends SLAM to a multi-agent setting, where robots share observations to collectively build a map and improve each other's pose estimates.

• Key Distinction: Single-agent SLAM relies solely on ego-motion and environmental features. Cooperative SLAM fuses inter-robot measurements.
• Example: A robot entering a new area can use range measurements to robots already mapped in that area to instantly correct its own drift.

The algorithmic combination of data from multiple sensors (e.g., IMU, camera, LiDAR, UWB) to create a coherent and accurate estimate of a system's state (pose, velocity). Cooperative localization is a form of distributed state estimation where the 'sensors' include other robots.

• Core Techniques: Kalman Filters (EKF, UKF), Particle Filters, and Factor Graphs are used to fuse proprioceptive (on-board) and exteroceptive (inter-robot) data.
• Challenge: Must account for correlated noise between robots, as their estimates become interdependent.

A system architecture where each robot makes decisions based on local information and communication with neighbors, without a central coordinator. Cooperative localization is often implemented in a decentralized fashion for robustness and scalability.

• Advantage: No single point of failure; the system degrades gracefully as robots are lost.
• Mechanism: Robots run local estimators and share only necessary data (e.g., beliefs or constraints) over a peer-to-peer network, converging on a consensus of the shared state.

Protocols that enable a team of distributed agents to agree on a common value (e.g., average position, leader identity, map data). In cooperative localization, consensus algorithms are used to ensure all robots converge to a mutually consistent understanding of their collective poses.

• Application: Robots may run a consensus on information algorithm to fuse their local estimates of a landmark's position or another robot's state.
• Example: The Consensus Kalman Filter allows robots to reach agreement on the global state estimate using only local communication.

The problem of planning collision-free paths for multiple agents from start to goal locations in a shared environment. Accurate cooperative localization provides the essential, high-fidelity positional data required for MAPF algorithms to function safely and efficiently.

• Dependency: MAPF plans (e.g., from Conflict-Based Search) assume agents know their positions relative to a shared map. Localization error can cause plan execution failures and collisions.
• Integration: Some advanced systems close the loop, using planned future paths as a prior to improve localization.

The graph defining which robots in a team can directly exchange information. The topology (e.g., fully connected, ring, dynamic ad-hoc) fundamentally limits the performance and convergence rate of cooperative localization algorithms.

• Impact: Sparse or intermittent links can lead to information islands and inconsistent state estimates across the team.
• Design Consideration: Algorithms must be robust to changing topologies, as robots move and links break. Techniques like consensus over unreliable networks are critical.