PaperlayerView exact revision on arXiv

Source-linked AI summary

Motion Planning and Control for Mobile Robot Navigation Using Machine Learning: a Survey

Xuesu Xiao, Bo Liu, Garrett Warnell, Peter Stone

arXiv:2011.13112v2cs.RO

TL;DR

Autonomous mobile robot navigation has been addressed by successful classical systems and newer machine-learning methods, but the two paradigms have lacked systematic comparison. This survey reviews learning-based motion planning and control within the classical navigation framework, classifies their relationships, and identifies opportunities to combine them for safer, more capable navigation.

  • Problem

    Classical and learning-based approaches address autonomous navigation, but their relationships and performance had not been systematically compared.

  • Method

    The survey organizes machine-learning approaches by their scope within classical navigation systems and compares them across functional scope, performance, tasks, and input modalities.

  • Results

    The survey finds opportunities to advance navigation by combining classical and learning-based paradigms.

  • Takeaways & Limitations

    Hybrid systems can combine learning modules for adaptation with classical components that provide hard safety constraints and explainability.

  • Takeaways & Limitations

    Many learning-based approaches address simplified moving-goal navigation and therefore cannot ensure arrival at a specified goal.

Abstract

from arXiv · show

Moving in complex environments is an essential capability of intelligent mobile robots. Decades of research and engineering have been dedicated to developing sophisticated navigation systems to move mobile robots from one point to another. Despite their overall success, a recently emerging research thrust is devoted to developing machine learning techniques to address the same problem, based in large part on the success of deep learning. However, to date, there has not been much direct comparison between the classical and emerging paradigms to this problem. In this article, we survey recent works that apply machine learning for motion planning and control in mobile robot navigation, within the context of classical navigation systems. The surveyed works are classified into different categories, which delineate the relationship of the learning approaches to classical methods. Based on this classification, we identify common challenges and promising future directions.

1 INTRODUCTION

Autonomous mobile robot navigation traditionally uses hierarchical global planning and local control, while machine learning offers data-driven alternatives. This survey organizes and compares these paradigms to identify their relationships, strengths, weaknesses, challenges, and future directions.

  • 1 INTRODUCTION: Classical navigation commonly combines global path planning with local motion control across perception, estimation, planning, and control.This hierarchical paradigm has enabled navigation on diverse mobile robot platforms.
  • 1 INTRODUCTION: End-to-end learning maps perceptual inputs directly to motion commands, bypassing symbolic intermediate representations and human-engineered navigation structure.Such systems can navigate without traditional symbolic, rule-based knowledge or engineering design.
  • 1 INTRODUCTION: Learning-based navigation has often required extensive data, lacked safety guarantees and explainability, and remained concentrated in simulation or simple real-world demonstrations.Collision causes may remain difficult to identify even after training on hundreds of thousands of instances.
  • 1 INTRODUCTION: Classical and learning-based navigation pursue the same problem but had not been systematically compared across largely separate research communities.The survey addresses this gap by examining learning approaches in the context of classical navigation.
  • 1 INTRODUCTION: The survey categorizes learning methods by their role in the navigation pipeline and compares them with classical approaches across functional scope and navigation performance.Its organization covers full-pipeline replacement, subsystem replacement, component improvement, performance categories, application domains, input modalities, and future directions.

2 CLASSICAL MOBILE ROBOT NAVIGATION

Classical mobile robot navigation decomposes motion-command generation into hierarchical global and local stages. Global planning produces a coarse route from a world representation, while local planning uses detailed nearby information and robot models to generate feasible commands.

  • 2 CLASSICAL MOBILE ROBOT NAVIGATION: The navigation problem generates motion-command sequences that move a robot from start location s to goal location g while minimizing a cost function J.The cost may encode motion constraints, obstacle avoidance, path or time, social compliance, stability, and other domain measures.
  • 2 CLASSICAL MOBILE ROBOT NAVIGATION: Classical systems decompose large-space navigation hierarchically because fine-grained long-horizon motion planning is computationally infeasible.Global and local stages divide the overall cost and reasoning burden.
  • 2.1 Global Planning: Global planning uses a coarse representation and evaluation criterion to produce a waypoint path toward the final goal, often treating the robot as a point mass.Representations include graphs, sampling-based structures, occupancy grids, and costmaps.
  • 2.2 Local Planning: Local planning receives a nearby goal or short global-path segment and generates kinodynamically feasible commands using a precise local representation and robot model.It reasons mainly about the robot’s immediate vicinity while optimizing objectives such as goal distance and obstacle clearance.
  • 2.2 Local Planning: The classical pipeline combines a global goal and representation with perceptual input to build a finer local representation and compute current motion commands.Figure 1 depicts global paths in blue, LiDAR returns in red, local representation in cyan, and commands in green.

3 SCOPE OF LEARNING FOR NAVIGATION

Machine learning has been proposed to reduce the engineering effort required by conventional navigation systems. The survey classifies learning approaches by how much of the classical navigation pipeline they replace or improve.

  • 3 SCOPE OF LEARNING FOR NAVIGATION: Machine learning navigation methods are motivated partly by the extensive engineering effort required to deploy conventional systems reliably in the real world.Classical systems often require substantial human tuning and engineering before deployment.
  • 3 SCOPE OF LEARNING FOR NAVIGATION: End-to-end approaches learn to map raw perceptual information directly to motion commands without explicit intermediate processing steps.They may drive the robot toward a specified goal or continuously forward.
  • 3 SCOPE OF LEARNING FOR NAVIGATION: The survey divides learning methods into replacing the entire navigation stack, learning navigation subsystems, and improving components within the stack.This classification organizes methods according to their functional scope relative to classical navigation.

3.1 Learning the Entire Navigation Stack

Learning-based approaches can replace the entire classical navigation stack, mapping sensor information and goals directly to motion commands. These works include fixed-goal and moving-goal settings, but moving-goal approaches do not address navigation to specified destinations.

  • Fixed-Goal Navigation: End-to-end methods treat the navigation system as a black box that maps perceptual signals and goals to discrete or continuous motion commands.They avoid explicitly formulating classical subcomponents and relationships between them.
  • Fixed-Goal Navigation: Fixed-goal approaches organize the literature by sensory input, including geometric, non-geometric, hybrid, and exteroceptive information.Geometric sensors directly indicate obstacles and free space, whereas non-geometric sensors such as RGB cameras do not.
  • Fixed-Goal Navigation: Fixed-goal learned systems have demonstrated obstacle avoidance and operation with other robots while directly producing commands for a specified destination.The surveyed approaches include imitation learning and reinforcement learning methods using inputs such as LiDAR, imagery, and goal locations.
  • Moving-Goal Navigation: Moving-goal approaches steer toward a goal that remains in front of the robot, enabling behaviors such as lane keeping and obstacle avoidance but not arrival at a specified goal.Their limited scope constrains the learning task and is conjectured to contribute to greater success in proof-of-concept systems.
  • Moving-Goal Navigation: Moving-goal end-to-end systems are limited in applicability to real navigation problems compared with fixed-goal navigation.They address forward movement and reactive behaviors rather than navigation to a specific destination.

3.2 Learning Navigation Subsystems

Learning navigation subsystems retains parts of the classical architecture while targeting specific planning functions. The surveyed subsystem work focuses predominantly on local planning across ground, aerial, and marine domains.

  • Subsystem Learning: Subsystem-learning approaches preserve other navigation subsystems as classical components while addressing more specific subproblems.This narrower scope allows learning methods to target selected parts of the classical architecture.
  • Local Planning: Most subsystem approaches replace the local planner, using a classical global planner to provide a local goal alongside current perceptual information.Learned modules then produce continuous motion control or local planning behavior.
  • Ground Navigation: Ground-navigation examples learn local behavior for intention-based control, reinforcement-learning local control, and socially aware obstacle avoidance.These systems combine local goals from A* or PRM with perception and learned policies.
  • Aerial Navigation: Aerial-navigation studies use imitation learning, reinforcement learning, or model-based reinforcement learning to learn local flight planning and control.One real-drone approach learned both the model and controller, using external motion-capture assistance.
  • Marine Navigation: Marine navigation work applies deep reinforcement learning to map surrounding-ship states to rudder commands while jointly addressing path following and collision avoidance.The unified reward incorporates the global planner’s path-following objective and collision avoidance.

3.3 Learning Individual Components

Learning individual components keeps the classical navigation pipeline while improving world representations or tuning planner parameters. These approaches seek to reduce manual design effort without replacing the full stack.

  • Individual Components: Component-learning approaches retain the classical navigation structure and target world representation or planner-parameter tuning.World representation bridges perception and planning in the classical stack.
  • World Representation: Learned world representations encode social constraints, navigation style, or terrain preferences for use by classical planners.Examples learn local cost functions or costmaps from demonstrations and then retain classical global or local planning modules.
  • Planner Parameters: Planner-parameter learning replaces extensive human expertise and labor with learned tuning of quantities such as inflation radius, sampling rate, and optimization weights.Examples learn parameters for GPMP2, DWA, or adaptive planner policies.
  • Planner Parameters: Adaptive planner policies interface with classical motion planners through hyper-parameters rather than issuing end-to-end motion commands.APPL supports teleoperated demonstrations, corrective interventions, evaluative feedback, and reinforcement learning.
  • Planner Parameters: Over 2x speedup in planning time was achieved by combining learned adaptive motion primitives with search algorithms on a 3 degree-of-freedom Reeds-Shepp problem.The approach was not implemented on physical mobile robots.
  • Individual Components: Hybrid component-learning methods aim to preserve classical safety, explainability, and reliability while improving weaknesses through learned representations or parameter tuning.The survey identifies this preservation-and-improvement combination as the principal benefit of these approaches.

3.4 Analysis

The survey finds that end-to-end learning dominates navigation research, while non-end-to-end approaches more consistently retain user-defined-goal capability. It therefore proposes hybrid classical-learning navigation as a promising direction.

  • Survey Observations: 43 out of the 74 surveyed papers focus on end-to-end techniques, despite their lack of proven reliable applicability in real-world scenarios.End-to-end learning obscures the functional blocks of the classical navigation pipeline.
  • Survey Observations: End-to-end navigation learns the entire stack jointly, but requires substantial training data and faces overfitting and explainability problems.Joint learning minimizes one end-to-end error rather than cascading component errors.
  • Survey Observations: One third of end-to-end approaches lack the ability to navigate to user-defined goals, instead targeting reactive behaviors such as lane keeping and obstacle avoidance.The survey attributes this difficulty possibly to poor generalization from trained goals to arbitrarily distant unseen goals.
  • Survey Observations: All 31 surveyed non-end-to-end approaches can navigate to user-defined goals while retaining desirable properties of existing navigation components.These approaches can also study learned improvements to other components.
  • Survey Observations: Of 13 subsystem approaches, 12 focus on local planning and only one considers global planning.The survey posits that local goals or path segments are easier to generalize than unseen faraway global goals.
  • Future Directions: The survey proposes a hybrid classical and learning navigation paradigm as a promising direction for future research.This proposal follows from the observed trade-offs between end-to-end learning and classical pipeline preservation.

4 COMPARISON TO CLASSICAL APPROACHES

The survey classifies learning-based navigation by how it relates to classical systems, then compares reported performance across duplication, alternate-interface, and improvement approaches. Evidence is mixed: many methods were untested against classical baselines, while selected hybrid, imitation-learning, and tuned systems matched or improved on classical approaches.

  • 4 COMPARISON TO CLASSICAL APPROACHES: Learning-based navigation is divided into methods duplicating classical input-output behavior, using alternate interfaces or constraints, and explicitly seeking improvement.These categories organize the surveyed literature according to its relationship with classical navigation systems.
  • 4 COMPARISON TO CLASSICAL APPROACHES: Several early end-to-end systems were not compared with classical navigation, while Pfeiffer’s approach achieved similar performance and Zhou’s was always outperformed by classical SLAM.The comparisons concerned fixed-goal, collision-free navigation using LiDAR input.
  • 4 COMPARISON TO CLASSICAL APPROACHES: Successor features and transfer learning produced task reward that outperformed DQN and rivaled classical A*, although reward may not reflect navigation performance.The reported result followed 8 hours of real experience, and video evidence still appeared inferior to classical approaches.
  • 4 COMPARISON TO CLASSICAL APPROACHES: Expert-demonstration pre-training reduced RL training time by a factor of five to reach at least the same performance, while the authors did not recommend replacing classical global planning.A PID controller was also reported to accelerate training and outperform standard DDPG, but that comparison was not against a classical planner.
  • 4 COMPARISON TO CLASSICAL APPROACHES: Learning achieved similar performance to MPC in one study and matched an MPC expert despite using lower-cost sensors in another.The latter system used a camera and wheel-speed sensor instead of the expert’s IMU and GPS.
  • 4 COMPARISON TO CLASSICAL APPROACHES: Automatic tuning, learned complements, and learned path costs target weaknesses of hand-tuned classical systems, including tedious trial-and-error parameterization.PRM-AutoRL was claimed to outperform PRM-DWA, but the classical approach was not similarly fine-tuned; LfH exceeded default and fine-tuned DWA in constrained spaces.

4.2 Learning beyond Classical Navigation

Learning beyond classical navigation addresses terrain-dependent behavior and social interaction, where hand-crafted costs and sequential effects are difficult to model. The surveyed approaches learn costs, policies, or controllers from demonstrations, experience, and interactions.

  • 4.2 Learning beyond Classical Navigation: Beyond classical geometric navigation, learning approaches target non-geometric terrain behavior and interactive social navigation with other robots or humans.These behaviors add physical-terrain or multi-agent considerations to goal-oriented navigation.
  • 4.2.1 Terrain-Based Navigation: Terrain-related navigation costs are difficult to hand-craft, motivating representation learning, imitation learning, and inverse reinforcement learning from demonstrations and experience.Applications include outdoor terrain-specific motion and human-like route preferences such as staying near grass or avoiding open areas.
  • 4.2.2 Social Navigation: Social navigation is difficult for classical methods because social cost functions are elusive and actions produce compounding sequential interactions among agents.These challenges motivate learned costs and reinforcement-learning controllers.
  • 4.2.2 Social Navigation: Social-navigation studies learned behavior through imitation, inverse reinforcement learning, inferred costs, human-motion data, and reinforcement-learning controllers.These methods included GAIL, RGB-D and global cost functions, surveillance-derived costmaps, and controllers incorporating social norms.
  • 4.2.2 Social Navigation: Learning-based methods outperformed interaction-unaware multi-robot approaches such as ORCA in reported social-navigation comparisons.CADRL achieved a 26% improvement in time to reach the goal over ORCA in simulation, while other methods improved success rate, extra time, and travel speed.

4.3 Analysis

The survey finds that learning-based navigation usually matches classical capabilities rather than surpassing them, while also enabling terrain-based and social navigation that classical techniques have struggled to achieve.

  • Most learning-based approaches solve classical navigation, but very few improve upon classical techniques.The survey characterizes learning-based navigation as mainly replicating capabilities already achievable with classical methods.
  • Learning-based methods have enabled relatively new capabilities such as terrain-based and social navigation.These capabilities have proven difficult to achieve with classical techniques.
  • 46 of 74 surveyed papers addressed the classical navigation problem, but only eight demonstrated improved performance over classical solutions.The other 38 generally operated in relatively simple environments and mostly lacked comparisons with classical approaches.

5 OTHER TAXONOMIES

The survey organizes prior work by six navigational tasks and four sensor-input modalities, showing that goal-oriented navigation with obstacle avoidance remains dominant while learning broadens navigation capabilities and sensing choices.

  • Navigational Tasks: The six task categories are Waypoint Navigation, Obstacle Avoidance, Waypoint Navigation + Obstacle Avoidance, Stylistic Navigation, Multi-Robot Navigation, and Exploration.All reviewed papers are assigned task categories in Figure 6.
  • Waypoint Navigation: Waypoint navigation moves a robot to a specified goal without considering additional constraints such as obstacle avoidance.The survey describes it as a fundamental building block, with relatively few works mainly serving as learning proof-of-concepts.
  • Waypoint Navigation + Obstacle Avoidance: Waypoint navigation with obstacle avoidance most closely matches classical navigation and requires reaching a specified goal while avoiding collisions.The majority of learning-based motion-planning and control approaches fall into this category.
  • Stylistic Navigation: Stylistic navigation generates behaviors that differ systematically from minimum-cost, collision-free motion, often using environmental features beyond geometric obstacles.Such tasks are difficult to define through classical static cost representations and have mostly been enabled by imitation or reinforcement learning.
  • Multi-Robot Navigation and Exploration: Multi-robot navigation accounts for other navigating agents, while exploration seeks to maximize environmental coverage for mapping or surveillance.Learning methods use experiential data for multi-robot policies and learned mappings or predictions for unknown space.
  • Input Modalities: Reviewed approaches use geometric sensors, RGB cameras, RGB plus geometry, or exteroceptive inputs.RGB-based learning methods often use monocular vision, while combined inputs pair geometric obstacle avoidance with visual semantic information; exteroceptive methods typically assume controlled environments and perfect sensing.
  • Summary: Goal-oriented navigation with obstacle avoidance remains the largest focus, while learning also targets stylistic and multi-robot navigation.The survey presents these taxonomies as an overview of learned navigation tasks and sensor modalities.

6 ANALYSIS AND FUTURE DIRECTIONS

The survey analyzes learning-based navigation across pipeline scope, classical comparison, task goals, and their cross-dimensional relationship. Most work remains end-to-end or component-level, usually addresses classical navigation without demonstrated superiority, while a smaller body extends navigation toward social and terrain-based capabilities.

  • Recap: Among the 31 non-end-to-end papers, 13 apply learning to a navigation subsystem and 18 to individual classical-system components.Subsystem studies include one global-planning and 12 local-planning papers; component studies include 14 representation and four parameter papers.
  • Comparison with Classical Approaches: 46 of 74 papers address classical navigation, but 38 of those were tested in simple environments without outperforming or often comparing against classical approaches.Only eight of the 46 were compared with classical approaches and demonstrated some improvement.
  • Beyond Classical Navigation: The remaining 28 papers pursue capabilities beyond classical navigation, including social navigation in 24 papers and terrain-based navigation in four.These approaches extend the goals considered by classical autonomous navigation.
  • Cross-Dimensional Analysis: Social-navigation approaches cluster at broad end-to-end or narrow component-level scopes, with only two intermediate subsystem approaches.The component-level works focus on learning cost-function representations, while terrain-based approaches span end-to-end, subsystem, and component categories.

COMPARISON

The survey compares machine learning approaches with classical hierarchical navigation and finds that narrower learning scopes generally correspond to better reported outcomes. It recommends subsystem- or component-level learning, especially in hybrid systems that retain classical safety and explainability advantages.

  • Comparison: Narrower machine learning scopes generally yield better results than broader applications across the classical navigation problem.The survey reports a correlation between learning scope and outcomes, with more limited scopes appearing to perform better.
  • Comparison: End-to-end approaches have not demonstrated improvement over classical navigation systems, while three subsystem-level works reported improved performance.End-to-end methods fall into Duplication with Standard or Alternative Interfaces and Constraints categories; subsystem-level methods span all categories, including Improvement.
  • Recommendations: The survey recommends applying machine learning at the subsystem or component level as the current best practice.This recommendation is based primarily on the cross-dimensional correlation between learning scope and overall system performance.
  • Recommendations: Hybrid systems can use learning modules to adapt to new environments while classical components enforce hard safety constraints.Classical navigation components also provide safety guarantees and explainability that machine-learning-only methods typically lack.
  • Future Directions: Reactive local planning and continual improvement from deployment experience are identified as promising directions for future machine-learning research.Only one of the 74 surveyed papers explicitly pursued continual onboard improvement using actual deployment experience.

7 CONCLUSIONS

The article organizes machine-learning navigation research in relation to the classical mobile-robot navigation pipeline and identifies opportunities to combine both paradigms. Such combinations may support faster, smoother, safer, and more reliable navigation in more constrained and diverse environments.

  • 7 CONCLUSIONS: The review classifies machine-learning approaches according to their relationships with classical navigation literature and the structured navigation pipeline.Its organization covers the learning methods' scope within structured systems and other relationships to classical navigation.
  • 7 CONCLUSIONS: Combining classical and learning-based navigation paradigms remains an opportunity for advancing the state of the art.The paper frames this opportunity after noting that the two research traditions have developed largely separately.
  • 7 CONCLUSIONS: The expected long-term goal is navigation that is quick, smooth, safe, and reliable in more constrained, diverse, and challenging environments.This expectation is stated as a prospective outcome of combining the two paradigms.
Loading 2011.13112v2…