11.3 Real-time decision making and adaptive mission planning

Real-time decision making is crucial for underwater robots navigating unpredictable environments. It allows them to adapt on the fly, avoiding obstacles and optimizing performance based on current conditions. This capability is essential for mission success and safety in dynamic underwater settings.

Adaptive mission planning takes this further, enabling robots to modify objectives and behaviors as circumstances change. This flexibility helps robots respond to new discoveries, equipment issues, or environmental shifts, maximizing efficiency and effectiveness in achieving mission goals.

Real-time Decision Making in Underwater Environments

Importance of Real-time Decision Making

Top images from around the web for Importance of Real-time Decision Making

• 

  Underwater robot deployed to aid endangered right whales - The Current View original

  Is this image relevant?

• 

  Woods Hole Oceanographic Institute Archives - Universe Today View original

  Is this image relevant?

• 

  Frontiers | A Mirror-Based Active Vision System for Underwater Robots: From the Design to Active ... View original

  Is this image relevant?

• 

  Underwater robot deployed to aid endangered right whales - The Current View original

  Is this image relevant?

• 

  Woods Hole Oceanographic Institute Archives - Universe Today View original

  Is this image relevant?

1 of 3

Top images from around the web for Importance of Real-time Decision Making

• 

  Underwater robot deployed to aid endangered right whales - The Current View original

  Is this image relevant?

• 

  Woods Hole Oceanographic Institute Archives - Universe Today View original

  Is this image relevant?

• 

  Frontiers | A Mirror-Based Active Vision System for Underwater Robots: From the Design to Active ... View original

  Is this image relevant?

• 

  Underwater robot deployed to aid endangered right whales - The Current View original

  Is this image relevant?

• 

  Woods Hole Oceanographic Institute Archives - Universe Today View original

  Is this image relevant?

1 of 3

• Underwater environments are highly dynamic and unpredictable, with constantly changing conditions such as currents, turbidity, and obstacles that can impact the performance and safety of underwater robots
• Real-time decision making allows underwater robots to adapt to these dynamic conditions by processing sensor data and making autonomous decisions in real-time, without relying on human intervention or pre-programmed instructions
• The ability to make real-time decisions is critical for underwater robots to effectively navigate, avoid obstacles, conserve energy, and complete their missions in the face of uncertainty and changing circumstances
• Real-time decision making enables underwater robots to respond quickly to unexpected events or emergencies, such as equipment failures or sudden changes in environmental conditions, which can help prevent accidents and ensure the success of the mission
• By making decisions in real-time based on the most current and relevant information available, underwater robots can optimize their performance, efficiency, and effectiveness in achieving their objectives
  • Enables robots to dynamically adjust their trajectories, speeds, and power consumption based on real-time measurements of currents, depths, and other environmental factors (energy efficiency)
  • Allows robots to detect and avoid unexpected obstacles or hazards that may not have been present during pre-mission planning (collision avoidance)

Benefits of Real-time Decision Making

• Increased adaptability and resilience in the face of changing conditions and uncertainties
  • Robots can modify their behaviors and strategies on-the-fly to cope with unforeseen challenges or opportunities (adjusting sampling locations based on real-time sensor readings)
• Enhanced safety and reliability of underwater operations
  • Real-time monitoring and decision making can help prevent accidents, collisions, and equipment failures by enabling robots to take proactive measures to mitigate risks (emergency surfacing or return to base in case of critical system faults)
• Improved efficiency and effectiveness in achieving mission objectives
  • Real-time optimization of robot actions based on current conditions can lead to faster, more accurate, and more comprehensive data collection and task execution (dynamically prioritizing survey areas based on real-time seafloor mapping data)
• Reduced reliance on human operators and pre-programmed instructions
  • Autonomous decision making capabilities allow robots to operate independently for longer periods of time and in more challenging environments, without the need for constant human supervision or intervention (long-duration underwater glider missions)

Factors Influencing Underwater Robotics Decisions

Sensor Data Quality and Reliability

• The quality and reliability of sensor data is a critical factor in real-time decision making for underwater robots, as inaccurate or incomplete data can lead to poor decisions and potentially catastrophic consequences
  • Noisy or biased sensor readings can cause robots to misinterpret their environment and make incorrect decisions (false positives in object detection leading to unnecessary avoidance maneuvers)
  • Sensor failures or malfunctions can deprive robots of essential information needed for decision making, leading to increased uncertainty and risk (loss of depth sensor data causing uncontrolled ascent or descent)
• Strategies for improving sensor data quality and reliability include:
  • Redundancy and diversity in sensor suites to provide multiple sources of information and cross-validation of measurements (using both sonar and optical sensors for obstacle detection)
  • Sensor fusion and filtering techniques to combine data from multiple sensors and reduce noise and uncertainties (Kalman filtering for state estimation)
  • Fault detection and isolation methods to identify and compensate for sensor failures or anomalies (using statistical tests or machine learning algorithms to detect outliers or inconsistencies in sensor data)

Computational and Power Constraints

• The computational power and processing speed of the robot's onboard computer systems can limit the complexity and sophistication of real-time decision making algorithms that can be implemented
  • Complex algorithms for perception, planning, and control may require significant computational resources that exceed the capabilities of embedded processors or low-power systems (deep learning models for image segmentation and classification)
  • Real-time constraints may necessitate the use of simplified or approximate algorithms that can generate decisions within the available time budget (heuristic search methods instead of exhaustive optimization)
• The available energy resources and power consumption of the robot can constrain the duration and intensity of real-time decision making processes, as well as the range and capabilities of the robot's sensors and actuators
  • Limited battery capacity may require robots to prioritize essential decision making tasks and conserve energy by reducing the frequency or resolution of sensor measurements (adaptive sampling strategies based on energy availability)
  • Power-hungry sensors or actuators may need to be used sparingly or in alternation to extend the robot's operational time and range (switching between high-power active sonars and low-power passive hydrophones for target tracking)

Mission Objectives and Parameters

• The specific mission objectives and parameters, such as the desired level of autonomy, the acceptable level of risk, and the priority of different goals, can influence the decision making strategies employed by the robot
  • Missions requiring high levels of autonomy and adaptability may necessitate more sophisticated and flexible decision making algorithms that can handle a wide range of scenarios and uncertainties (multi-objective optimization for balancing exploration and exploitation in search and rescue operations)
  • Missions with strict safety or performance requirements may impose more conservative decision making strategies that prioritize risk avoidance and reliability over efficiency or speed (using larger safety margins and slower speeds in cluttered or hazardous environments)
  • Missions with multiple, potentially conflicting objectives may require decision making algorithms that can make trade-offs and compromises based on the relative importance and feasibility of different goals (prioritizing sample collection over area coverage in time-limited scientific surveys)

Environmental Uncertainties and Variability

• The inherent uncertainties and variability of underwater environments, such as the presence of unknown obstacles, changing currents, and fluctuating visibility, can make real-time decision making more challenging and require more robust and adaptive algorithms
  • Unknown or partially observable environmental features may require probabilistic or stochastic decision making approaches that can reason about uncertainties and update beliefs based on new observations (using occupancy grid maps or belief maps for navigation in unstructured environments)
  • Dynamic or unpredictable environmental conditions may necessitate reactive or model-predictive decision making strategies that can anticipate and adapt to changes in real-time (using online learning or adaptive control methods to estimate and compensate for time-varying currents or disturbances)
  • Harsh or extreme environmental conditions may require decision making algorithms that are robust to sensor noise, communication dropouts, and other sources of uncertainty or failure (using redundant or fail-safe control architectures for operation in deep or turbid waters)

Adaptive Mission Planning for Dynamic Environments

Adaptive Mission Planning Techniques

• Adaptive mission planning involves the dynamic modification of the robot's objectives, trajectories, and behaviors based on real-time feedback and changing circumstances, rather than following a fixed, pre-determined plan
  • Enables robots to respond to new opportunities or challenges that arise during the mission, such as the detection of unexpected targets or the failure of certain subsystems (dynamically re-planning the mission to investigate a newly discovered hydrothermal vent or to return to base after a critical component failure)
  • Allows robots to optimize their performance and resource utilization based on the actual conditions encountered in the field, rather than relying on potentially inaccurate or outdated prior information (adjusting the spacing and duration of survey transects based on real-time measurements of seafloor complexity and variability)
• Probabilistic planning techniques, such as Markov Decision Processes (MDPs) and Partially Observable Markov Decision Processes (POMDPs), can be used to model the uncertainties and rewards associated with different actions and states, and to generate optimal policies that maximize the expected utility of the mission
  • MDPs represent the environment as a set of discrete states and actions, with transition probabilities and rewards associated with each state-action pair, and can be solved using dynamic programming or reinforcement learning methods to find the optimal policy (using value iteration to compute the optimal navigation policy for a grid-based map of the environment)
  • POMDPs extend MDPs to situations where the robot has incomplete or noisy observations of the environment, and maintain a belief state representing the probability distribution over the possible states, which is updated using Bayesian inference based on the robot's actions and observations (using point-based value iteration to plan inspection tasks in the presence of sensor uncertainty and occlusions)

Contingency Planning and Fault Tolerance

• Contingency planning strategies involve the identification of potential failure modes and the development of alternative courses of action to mitigate risks and ensure the successful completion of the mission, even in the face of unexpected events or degraded performance
  • Fault tree analysis and failure mode and effects analysis (FMEA) can be used to systematically identify and prioritize the potential failure scenarios and their consequences, and to design appropriate contingency plans and safeguards (using redundant or backup systems for critical components, such as multiple navigation sensors or communication links)
  • Graceful degradation and reconfiguration techniques can be employed to enable the robot to continue operating with reduced capabilities or in a safe mode when certain subsystems fail or become unavailable (using thruster allocation algorithms to maintain stable motion control in the event of individual thruster failures)
  • Adaptive mission planning algorithms can incorporate contingency plans and fault tolerance mechanisms into the decision making process, by considering the likelihood and impact of different failure scenarios and selecting actions that minimize the overall risk and maximize the expected mission success (using chance-constrained programming or risk-aware planning methods to generate robust mission plans that can handle uncertainties and faults)

Hierarchical and Multi-robot Planning

• Hierarchical planning approaches can be used to decompose complex missions into smaller, more manageable sub-tasks, and to enable the robot to adapt its behavior at different levels of abstraction based on the changing requirements and constraints of the mission
  • High-level mission planning can focus on the overall goals, priorities, and constraints of the mission, and generate a sequence of subgoals or waypoints that guide the robot's behavior at a coarse resolution (using a symbolic planner to generate a high-level mission script that specifies the main survey areas, sampling locations, and communication checkpoints)
  • Low-level motion planning and control can focus on the detailed execution of each subgoal or waypoint, and generate fine-grained trajectories and control commands that take into account the local environmental conditions and the robot's dynamics and constraints (using a sampling-based planner and a feedback controller to generate smooth and feasible paths between waypoints while avoiding obstacles and currents)
  • Hierarchical planning allows for a modular and scalable approach to mission planning, where different planning algorithms and models can be used at each level of the hierarchy, and where the higher-level plans can be adapted or re-planned based on the feedback and performance of the lower-level plans (using a mission executive to monitor the progress and completion of each subgoal, and to trigger re-planning or contingency actions when necessary)
• Collaborative planning strategies can be developed to enable multiple robots to work together and adapt their individual behaviors to achieve common goals, while taking into account the capabilities, limitations, and actions of other robots in the team
  • Centralized planning approaches rely on a single master node or ground station to generate and coordinate the plans for all the robots in the team, based on a global view of the environment and the mission objectives (using a mixed-integer linear programming formulation to optimize the task allocation and trajectories for a fleet of autonomous underwater vehicles in a seafloor mapping mission)
  • Decentralized planning approaches allow each robot to generate its own plans based on local information and communication with nearby robots, using distributed algorithms for consensus, task allocation, and collision avoidance (using a market-based bidding protocol for dynamic task assignment and a distributed model predictive control scheme for coordinated motion planning in a multi-robot underwater search mission)
  • Hybrid planning approaches combine elements of centralized and decentralized planning, by using a hierarchical or leader-follower structure where some robots act as coordinators or supervisors for groups of subordinate robots, and where the planning and decision making responsibilities are distributed across different levels of the hierarchy (using a cluster-based planning architecture for cooperative underwater surveillance and intervention missions, where cluster heads coordinate the actions of their member robots and communicate with other cluster heads to achieve global objectives)

Real-time Decision Making Algorithms for Underwater Robots

Perception and Sensor Fusion

• Real-time decision making algorithms for underwater robots typically involve a combination of sensing, perception, planning, and control modules that work together to enable the robot to make intelligent decisions and take appropriate actions in response to its environment
• Sensor fusion techniques, such as Kalman filtering and particle filtering, can be used to combine data from multiple sensors and to estimate the robot's state and the state of its environment, while dealing with noise, uncertainty, and inconsistencies in the sensor measurements
  • Kalman filters use a linear dynamical model and Gaussian noise assumptions to recursively estimate the optimal state estimate and its uncertainty from noisy sensor measurements and control inputs, and can be extended to nonlinear systems using linearization techniques such as the extended Kalman filter (EKF) or the unscented Kalman filter (UKF) (using an EKF to fuse data from a depth sensor, a Doppler velocity log, and an inertial measurement unit for underwater localization and navigation)
  • Particle filters use a non-parametric representation of the state distribution as a set of weighted samples, and can handle non-linear and non-Gaussian systems by approximating the posterior distribution using importance sampling and resampling techniques (using a particle filter to track the position and velocity of an underwater target from noisy sonar measurements and a motion model)
• Object detection and recognition algorithms, such as convolutional neural networks and template matching, can be employed to identify and classify relevant features and obstacles in the robot's surroundings, and to provide semantic information for decision making and planning
  • Convolutional neural networks (CNNs) use a hierarchical structure of convolutional, pooling, and fully connected layers to learn discriminative features and classify images into different categories, and can be trained on large datasets of labeled underwater images to detect and recognize objects of interest (using a CNN to detect and classify different types of marine debris from underwater camera images in real-time)
  • Template matching techniques use predefined patterns or features to search for similar regions in an image, and can be used for fast and efficient detection of known objects or landmarks in the environment (using normalized cross-correlation to detect and track a specific type of underwater pipeline from side-scan sonar images)

Planning and Control

• Path planning and obstacle avoidance algorithms, such as A* search, Rapidly-exploring Random Trees (RRTs), and potential field methods, can be used to generate safe and efficient trajectories for the robot to follow, while avoiding collisions with obstacles and optimizing various performance criteria
  • A* search is a heuristic-based graph search algorithm that finds the optimal path between a start and a goal node in a discretized environment, using an admissible heuristic function to estimate the cost-to-go and prioritize the expansion of promising nodes (using A* search to plan the shortest collision-free path for an autonomous underwater vehicle in a grid-based map of the environment)
  • RRTs are sampling-based motion planning algorithms that incrementally build a tree of feasible trajectories by randomly sampling points in the configuration space and attempting to connect them to the nearest node in the tree, while checking for collisions and constraints (using RRTs to plan dynamic and kinodynamic trajectories for an underwater manipulator arm in a cluttered environment)
  • Potential field methods use artificial potential functions to represent the attractiveness of the goal and the repulsiveness of obstacles, and generate motion commands that follow the gradient of the potential field to reach the goal while avoiding obstacles (using a harmonic potential field to plan smooth and stable trajectories for an underwater glider in the presence of ocean currents and eddies)
• Control algorithms, such as proportional-integral-derivative (PID) control, model predictive control (MPC), and adaptive control, can be implemented to execute the planned actions and to maintain the robot's stability and performance in the face of disturbances and uncertainties
  • PID control is a simple and widely used feedback control technique that calculates the control signal as a weighted sum of the error, its integral, and its derivative, and can be tuned to achieve a desired balance between responsiveness, stability, and robustness (using a PID controller to regulate the depth and heading of an autonomous underwater vehicle using feedback from pressure and compass sensors)
  • MPC is an optimization-based control technique that repeatedly solves a constrained finite-horizon optimal control problem over a receding horizon, using a model of the system dynamics and constraints to predict the future behavior and optimize the control inputs (using MPC to control the trajectory and energy consumption of an underwater glider in the presence of time-varying currents and disturbances)
  • Adaptive control techniques, such as model reference adaptive control (MRAC) and self-tuning regulators (STR), can automatically adjust the control parameters or the model parameters based on the observed performance and the changing environment, to maintain the desired closed-loop behavior (using MRAC to control the position and orientation of an underwater robot in the presence of unknown hydrodynamic coefficients and time-varying payloads)

High-level Reasoning and Autonomy

• Decision trees, rule-based systems, and finite state machines can be used to encode the robot's high-level behaviors and to enable it to switch between different modes of operation based on the current situation and the mission requirements
  • Decision trees are hierarchical structures that represent a set of decision rules and their consequences, and can be used to classify the current situation and select the appropriate behavior based on a series of tests on the sensor data and the mission parameters (using a decision tree to select between different survey patterns and sampling strategies based on the seafloor type, the water depth, and the scientific objectives)
  • Rule-based systems use a set of if-then rules to encode the domain knowledge and the expert reasoning, and can be used to trigger specific actions or behaviors based on the satisfaction of certain conditions or the occurrence of certain events (using a rule-based system to manage the power and communication systems of an underwater sensor network, based on rules for battery charging, data transmission, and fault detection)
  • Finite state machines are graph-based models that represent the system as a set of discrete states and transitions between them, and can