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

Classical mobile-robot navigation is well established, but direct comparison with emerging machine-learning approaches has been limited. This survey reviews and classifies learning-based planning and control in the context of classical navigation, concluding that combining the paradigms offers opportunities to advance navigation.

  • Problem

    Direct comparison between classical navigation systems and emerging machine-learning approaches has been limited, despite both addressing mobile-robot navigation.

    Source: §1 · p. 2

  • Method

    The survey qualitatively compares recent learning-based motion-planning and control approaches with classical navigation, organizing them by their role and scope within navigation systems.

    Source: §1 · p. 2, §7 · p. 28

  • Results

    The survey identifies opportunities to advance mobile-robot navigation by combining classical and learning-based approaches.

    Source: §7 · p. 29

  • Takeaways & Limitations

    The authors expect this combination to support robots navigating safely and reliably through more constrained, diverse, and challenging environments.

    Source: §7 · p. 29

    Moving-goal learning approaches can support behaviors such as lane keeping and obstacle avoidance but cannot ensure arrival at a specified goal, limiting their applicability to real navigation problems.

    Source: §3.1.2 · pp. 9–10

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

The survey examines machine learning for mobile robot motion planning and control alongside classical navigation, addressing the lack of systematic comparison between these approaches. It categorizes learning methods by their relationship to the classical navigation pipeline and identifies challenges and future research directions.

Introduction

  • Classical hierarchical navigation has enabled deployment across mobile robot platforms, but remains less robust in highly constrained or diverse environments than human teleoperation.These systems typically combine global path planning with local motion control.Source: §1 · p. 1
  • End-to-end learning maps perceptual inputs directly to motion commands, but is data-hungry, lacks safety guarantees and explainability, and has mostly been tested in simulation or simple real-world settings.When a collision occurs, its cause may remain difficult to identify even after training on hundreds of thousands of examples, hindering debugging and improvement.Source: §1 · pp. 1–2
  • The survey qualitatively compares recent learning approaches with classical methods, organizing learning by its role in navigation and examining functional scope and navigation performance.It considers whether learning replaces navigation end to end or addresses an individual task, among other dimensions.Source: §1 · p. 2
  • The survey aims to highlight learning advances, organize methods by their role in navigation, discuss their strengths and weaknesses, and identify future research directions.It also groups learning work by whether it replaces the full navigation pipeline, replaces a subsystem, or improves existing components, then examines performance categories, application domains, input modalities, challenges, and future directions.Source: §1 · p. 2
  • Unlike prior surveys centered on learning methods, this survey relates learning work to the classical navigation pipeline and includes approaches that move robots in simulated or physical environments.It excludes, for example, gridworld agents, semantic mapping not used to move robots, and robotic manipulation; it aims to identify classical problems explored or still unexplored from a learning perspective and navigation problems classical methods have not solved.Source: §1 · p. 2

2 CLASSICAL MOBILE ROBOT NAVIGATION

Classical mobile robot navigation seeks a cost-minimizing sequence of motion commands from a start to a goal. Because long-horizon planning can be computationally infeasible, systems commonly divide it into global and local reasoning.

2 CLASSICAL MOBILE ROBOT NAVIGATION

  • Classical navigation generates a cost-minimizing action sequence from start to goal, decomposing large-horizon planning hierarchically into global and local reasoning.Actions depend on the robot, while the objective may account for motion constraints, obstacle avoidance, distance or time, and other domain-dependent measures.Source: §2 · p. 3

2.1 Global Planning

  • Global planning builds a coarse world representation and uses efficient search to produce waypoint paths, often treating the robot as a point mass and replanning when observations change.Representations may use sensor data and maps, with sparse structures such as PRMs and RRTs, or occupancy grids and costmaps; fine details may be left to local planning.Source: §2.1 · pp. 3–4

2.2 Local Planning

  • Local planners use precise nearby perception and robot models to generate kinodynamically feasible commands that pursue local goals while respecting constraints and avoiding obstacles.They optimize objectives such as distance to the local goal and obstacle clearance, outputting discrete commands or continuous velocities for low-level motor controllers.Source: §2.2 · p. 4

3 SCOPE OF LEARNING FOR NAVIGATION

Machine-learning navigation methods are surveyed as potential ways to reduce the engineering effort required by classical approaches. The section organizes this work by whether learning replaces the full navigation stack or targets particular subsystems or components.

3 SCOPE OF LEARNING FOR NAVIGATION

  • Machine learning is proposed as a potential way to reduce the extensive engineering effort classical navigation approaches require for reliable real-world deployment.The survey covers learning methods addressing problems within the classical mobile-robot navigation pipeline.Source: §3 · p. 5
  • Learning methods either map raw perceptual input directly to motion commands or augment or replace parts of the classical navigation pipeline.End-to-end methods may move toward a specified goal or continuously forward without explicit intermediate processing.Source: §3 · p. 5
  • The survey groups work into learning to replace the entire navigation stack, learning navigation subsystems, and learning components within the stack.These categories correspond to Sections 3.1, 3.2, and 3.3, respectively.Source: §3 · p. 5
  • Figure 2 distinguishes full-stack methods by goal type and command type, and depicts learned global or local planning and individual components.Its cell shading indicates the number of surveyed papers in each category; Figure 3 illustrates the structure of Section 3.Source: §3 · p. 5

3.1 Learning the Entire Navigation Stack

End-to-end navigation learning commonly treats the pipeline as a black box mapping perception to motion commands. The survey distinguishes approaches that use a fixed destination from those that react to a moving goal.

3.1.1 Fixed-Goal Navigation

  • Fixed-goal approaches are grouped by whether they use geometric sensors, non-geometric sensors, or a combination of both.Examples include LiDAR, RGB imagery, and hybrid systems combining visual and LiDAR inputs.Source: §3.1.1 · pp. 5, 8
  • Fixed-goal geometric systems have used end-to-end policies trained by imitation or reinforcement learning, with some methods adding memory or exploration rewards.One approach imitates the ROS move base stack, while others use recurrent memory, neural mapping, or exploration-encouraging rewards.Source: §3.1.1 · p. 7
  • A case study cautions against fully replacing classical navigation with end-to-end learning, recommending map-based global planning for reliable navigation in unseen environments.The study examined end-to-end architectures and concluded that a classical global planner should be retained when possible.Source: §3.1.1 · p. 8
  • Fixed-goal learned systems can avoid obstacles, navigate among robots, account for human social norms, and explore unknown environments.These approaches take a fixed goal and sensor data as input and produce discrete or continuous motion commands.Source: §3.1.1 · p. 9

3.1.2 Moving-Goal Navigation

  • Moving-goal methods enable behaviors such as lane keeping and collision avoidance, but cannot ensure arrival at a specified destination.The survey suggests their narrower scope constrains the learning task and can help existing machine-learning methods succeed.Source: §3.1.2 · pp. 9–10
  • Moving-goal learning spans indoor and outdoor ground robots and UAVs, using demonstrations, imitation learning, or reinforcement learning to produce navigation behaviors.Examples include corridor collision avoidance, terrain adaptation, and vision-based UAV flight.Source: §3.1.2 · pp. 9–10

3.2 Learning Navigation Subsystems

Learning-based navigation can replace selected subsystems while retaining classical components, allowing learning to target specific subproblems. Most work focuses on local planning, with one surveyed approach targeting global planning.

3.2 Learning Navigation Subsystems

  • Subsystem learning preserves other classical navigation components while targeting specific subproblems; most surveyed work addresses local planning, and one work addresses global planning.The learned subsystem interfaces with another subsystem implemented classically.Source: §3.2 · p. 10

3.2.1 Learning Global Planning

  • Yao, et al. [107] used Group-Navi GAN as a global planner to help a robot join social-group flow, while state estimation and collision avoidance remained classical.The network tracked social groups and generated motion plans that provided local goals to the local planner in densely populated environments.Source: §3.2.1 · p. 10

3.2.2 Learning Local Planning

  • Local-planning methods typically use a classical global planner’s local goal and current perception to produce continuous motion control, replacing only the local-planner subsystem.The surveyed works address only the local-planning part of the traditional navigation stack.Source: §3.2.2 · pp. 10–11
  • Ground-navigation approaches learned local control from global-planner goals, using imitation learning, reinforcement learning, PID-assisted training, or automated reward and architecture search.Intention-Net used a path-image intention and perception to produce speed and steering commands; PRM-RL paired PRM long-range planning with an RL short-range policy, and PRM-AutoRL searched reward functions and network architectures. Xie, et al. used a classical PID controller to kick-start local-controller RL training.Source: §3.2.2 · pp. 10–11
  • Other ground methods targeted social navigation, terrain-aware action learning, safe agile maneuvers, planner complementarity, and off-road control.Pokle, et al. used attention to approach goals, avoid obstacles, and respect pedestrian space. BADGR learned from physical interaction without simulation or human supervision, but training could cause catastrophic failures requiring manual intervention. Learning from Hallucination, with Sober Deployment and learned hallucination, learned local planners safely in free space for constrained environments. Lifelong Navigation trained a neural planner from DWA data to complement DWA when its behavior was suboptimal, using GEM to avoid catastrophic forgetting; Xiao, et al. used inertial terrain signatures to learn wheel-terrain interactions for high-speed off-road navigation.Source: §3.2.2 · p. 11
  • Aerial methods used imitation learning followed by reinforcement learning for narrow-gap UAV flight, and model-based reinforcement learning for obstacle-free drone point-to-point flight.Lin, et al. first learned a local-planning policy by imitation learning and then improved it with reinforcement learning. Becker, et al. [1] learned both model and controller and used an external motion-capture system.Source: §3.2.2 · p. 11
  • For marine navigation, Zhao and Roh used DRL to map fully observable surrounding-ship states to three discrete rudder commands, with a unified reward for path following and collision avoidance.Source: §3.2.2 · p. 11

3.3 Learning Individual Components

This section surveys learning methods that preserve the classical navigation pipeline by improving individual components, chiefly world representation and planner parameters.

3.3.1 World Representation

  • For social navigation, learned costs and representations encode human norms or trajectories while classical planners, including A*, Theta*, RRT, and POSQ, plan routes,,,,,,,,,,.Approaches include learning local costs from RGB-D demonstrations, modeling social norms from observed navigation, and using learned maps or path predictions to guide planners,.Source: §3.3.1 · pp. 11–12
  • Learned world representations also support navigation styles, collision-aware planning, and goal seeking in unknown environments,,.Methods learn local costmaps to imitate demonstrations, use novelty classification to inform collision costs, and estimate the cost-to-go from subgoals at known–unknown boundaries.Source: §3.3.1 · p. 12

3.3.2 Planner Parameters

  • Learning can tune existing planner parameters: GPMP2 learns obstacle covariance from demonstrations, while Predictive DWA uses ANFIS to adjust its optimization weights.Predictive DWA modifies its three weights individually using hand-engineered features and backpropagation.Source: §3.3.2 · p. 12
  • APPL learns planner parameters from human interaction and adjusts them at runtime, outperforming planners with fixed parameters tuned by human experts [104].Its modalities include teleoperated demonstrations (APPLD), corrective interventions (APPLI), evaluative feedback (APPLE), and reinforcement learning (APPLR); its policy changes planner hyperparameters rather than issuing end-to-end motion commands.Source: §3.3.2 · p. 12
  • Over 2x speedup in planning time was achieved by learned adaptive motion primitives combined with search algorithms in a 3-degree-of-freedom Reeds–Shepp navigation problem, without physical-robot implementation.The reported result concerns the planning problem rather than experiments on a physical mobile robot.Source: §3.3.2 · pp. 12–13

3.3 Learning Individual Components

  • These methods retain classical navigation components and their safety, explainability, and reliability while addressing their weaknesses through learning.Learning targets global or local planners and individual components, including world representation and planner parameters.Source: §3.3 · p. 11, §3.3.2 · p. 13

3.4 Analysis

The analysis contrasts end-to-end learning’s prevalence and tradeoffs with goal-capable non-end-to-end methods, and proposes hybrid classical-learning navigation as a promising research direction.

3.4 Analysis

  • 43 of 74 surveyed papers use end-to-end techniques, despite their lack of proven reliability in real-world scenarios.They simplify development by replacing hand-crafted components and may avoid cascading errors, but face high training-data demands, overfitting, and limited explainability.Source: §3.4 · p. 13
  • One third of end-to-end approaches lack user-defined goals, reflecting difficulty generalizing to unseen distant goals; Pfeiffer et al. argue they cannot replace map-based planners in large complex environments.These systems instead target reactive behaviors such as lane keeping and obstacle avoidance.Source: §3.4 · p. 13
  • All 31 surveyed non-end-to-end approaches can navigate to user-defined goals while retaining existing components and applying learning to improve cost functions, planner parameters, or overall performance.Source: §3.4 · p. 13
  • Of 13 subsystem-learning approaches, 12 focus on local planning and one on global planning, possibly because local goals generalize more readily than unseen distant goals.Source: §3.4 · pp. 13–14
  • The analysis proposes a hybrid classical and learning navigation paradigm as a promising direction for future research.Source: §3.4 · p. 14

4 COMPARISON TO CLASSICAL APPROACHES · 4.1 Learning for Navigation in the Classical Sense

The section compares learning-based navigation with classical systems, focusing on approaches that seek classical collision-free, goal-directed behavior from geometric perception and relatively accurate motion models. It groups these approaches by whether they duplicate standard systems, use alternate interfaces or constraints, or seek to outperform classical methods.

4.1 Learning for Navigation in the Classical Sense

  • Classical-sense learning is grouped into standard-interface duplication, alternate-interface or constraint use, and methods explicitly seeking to outperform classical navigation.The focus is collision-free motion toward a specified goal based on geometric perception and a relatively accurate motion model.Source: §4.1 · p. 14

4.1.1 Duplication with Standard Interfaces and Constraints

  • Among early fixed-goal LiDAR approaches, Pfeiffer, et al. achieved performance similar to classical systems, while Zhou, et al. was outperformed by SLAM-based navigation.Other early systems lacked classical comparisons: target-pursuit methods by Thrun and Zhu, et al., obstacle avoidance by Sergeant, et al., and moving-goal methods by Tai, et al. and Zhao and Roh.Source: §4.1.1 · pp. 14–15
  • Later duplication work often measured learning-specific gains rather than classical performance; modular RL beat Potential Field Method, but not necessarily stronger SLAM-based systems.Successor features rivaled classical A* on task reward after eight hours of real experience, though the authors’ video appeared inferior; curiosity-driven DRL improved on vanilla A3C, Neural SLAM beat stacked-LSTM A3C, and MACN’s exploration-path ratio to A* was never below one.Source: §4.1.1 · pp. 15–16
  • Learning-focused training improvements included expert pretraining that cut time to a target performance fivefold and a PID warm start that trained faster and matched ROS move base.Lin, et al. reported performance similar to MPC by combining IL and RL with full MoCap state information, possibly with lower thrust; Pfeiffer, et al. did not recommend replacing classical global planners.Source: §4.1.1 · p. 16

4.1.2 Duplication with Alternate Interfaces and Constraints

  • Visual learning can bypass brittle or costly engineered vision pipelines, but the surveyed visual-navigation approaches lacked experimental comparisons with classical counterparts.Examples include end-to-end image-to-action systems such as DAVE-2 [3] and conditional imitation learning for fixed-goal navigation by Codevilla, et al..Source: §4.1.2 · pp. 16–17
  • Learning can leverage platform data to handle complex dynamics and partial observability, but Faust, et al. compared only against a trivial straight-line planner.Approaches include UAV obstacle avoidance trained with full state but deployed under partial observation, RL for non-trivial ground and aerial dynamics, and obstacle-free learned flight dynamics [1].Source: §4.1.2 · p. 17
  • Low-cost sensing and hardware were explored through navigation with only ten LiDAR beams and an off-road vehicle that matched its better-equipped MPC expert.Pan, et al. trained from MPC demonstrations using high-precision IMU and GPS, then deployed with a camera and wheel-speed sensor.Source: §4.1.2 · pp. 17–18

4.1.3 Improvement

  • Improvement approaches learned planner parameters, motion primitives, complementary controllers, or path costs, with reported gains over specified classical baselines.APPL [104] improved over fixed default or manually tuned parameters; learned motion primitives reduced computation time. PRM-AutoRL was claimed to beat PRM-DWA without equivalent classical fine-tuning, while LfH, Lifelong Navigation, and Learned Subgoal Planner reported gains over their cited baselines.Source: §4.1.3 · p. 18

4.2 Learning beyond Classical Navigation

Classical navigation handles goal-oriented, geometric, collision-free movement, but more complex behaviors have motivated learning-based research on non-geometric terrain navigation and social navigation among other agents.

4.2.1 Terrain-Based Navigation

  • Learning-based terrain navigation can capture ground properties that are difficult to hand-code, using demonstrations, inverse reinforcement learning, physical interaction, or learned dynamics.Examples include learning to favor tall grass over uneven terrain and conditioning an inverse kinodynamics model on inertia embeddings for high-speed off-road navigation.Source: §4.2.1 · p. 19

4.2.2 Social Navigation

  • Social navigation is difficult to specify with static costs because social norms vary and agents’ actions create sequential interactions that affect later decisions.Human-aware costs may allow closer passage in crowded spaces or when people and robots move in the same direction.Source: §4.2.2 · p. 19
  • Social navigation methods learn costs or policies from demonstrations: GAIL learns human-like policies, while IRL and trajectory or graph methods model social costs and favorable behavior,,,,,.Other approaches infer costs from human motion prototypes, exploration data, or demonstrations, and use observed trajectories or behavior to shape planning,.Source: §4.2.2 · pp. 19–20
  • Reinforcement learning offers a way to model compounding social interactions, with controller designs addressing multi-agent collision avoidance, social passing norms, pedestrian interactions, and companion travel,,,,.Examples include GA3C-CADRL, which adds an LSTM and GPU to learn policies for complex sequential interactions, and roleplaying learning for companion travel.Source: §4.2.2 · p. 20
  • 26% improvement in time to reach the goal over ORCA was reported for CADRL, while other learned systems also outperformed ORCA and additional navigation baselines.The surveyed comparisons include RL-based approaches using learned velocities, decentralized collision avoidance, multi-sensor fusion, interaction models, hierarchical control, and attention-based planning.Source: §4.2.2 · p. 20

4.3 Analysis

The survey finds that learning-based navigation has mostly replicated capabilities already achievable with classical approaches, with few demonstrated improvements. Learning has also enabled newer capabilities, particularly social and terrain-based navigation, that are difficult for classical techniques.

4.3 Analysis

  • Learning-based methods have mostly replicated classical navigation, with few surveyed approaches demonstrating better performance.The authors advocate comparing learning and classical solutions, including training versus engineering effort and neural architecture search versus parameter tuning.Source: §4.3 · pp. 20–21
  • Of 46 papers addressing the classical navigation problem, only eight showed improved performance; the other 38 mostly used simple environments and lacked classical comparisons.The authors suggest the field initially focused on whether learning-based navigation was possible and on its potential to avoid extensive human engineering or leverage alternative interfaces and constraints.Source: §4.3 · p. 21
  • Of 28 papers pursuing capabilities beyond classical navigation, 24 focused on social navigation and four on terrain-based navigation.Social navigation may predominate because humans and pedestrian environments are more accessible to researchers than challenging terrain; these capabilities have been difficult to achieve classically.Source: §4.3 · p. 21

5 OTHER TAXONOMIES

The survey classifies its reviewed learning-based navigation papers by the task they address and by their sensor input modality. The taxonomies show goal-oriented navigation with obstacle avoidance as the largest task focus, alongside growing attention to interactive behaviors and visual inputs.

5.1 Navigational Tasks

  • The task taxonomy groups reviewed papers into six navigation tasks, while the input taxonomy distinguishes four sensor modalities.Figures 6 and 7 categorize the 74 papers within the survey’s stated scope of learning-based motion planning and control that actually moves mobile robots in their environments.Source: §5.1 · p. 21, §5 · p. 21, §5.2 · p. 23

5.1.1 Waypoint Navigation

  • Waypoint navigation moves to a specified goal without other constraints, while obstacle-avoidance methods address obstacles without navigating to a prespecified goal.Waypoint-only work is sparse and often proof-of-concept; hallway navigation represents obstacle avoidance, which includes all Moving-Goal Navigation approaches discussed in Section 3.1.2.Source: §5.1.1 · p. 21, §5.1.2 · p. 21

5.1.3 Waypoint Navigation + Obstacle Avoidance

  • Waypoint navigation with obstacle avoidance most closely matches classical navigation and contains the majority of surveyed learning-based approaches.It requires reaching a specified goal while avoiding collisions along the way.Source: §5.1.3 · p. 22

5.1.4 Stylistic Navigation

  • Stylistic navigation generates behaviors that differ from minimum-cost, collision-free motion, such as hallway-side preferences or terrain-dependent speed.Because these tasks are difficult to specify with classical static costs, they have mostly been enabled by IL and RL.Source: §5.1.4 · p. 22

5.1.5 Multi-Robot Navigation

  • Multi-robot navigation explicitly accounts for other agents, and learning methods use experiential data to find policies where hand-crafted classical heuristics become insufficient as agent numbers grow.Source: §5.1.5 · p. 22

5.1.6 Exploration

  • In pure exploration, the agent seeks environment coverage for tasks such as mapping or surveillance, using learning to memorize explored space or predict unknown space from training experience.Source: §5.1.6 · p. 22

5.2.1 Geometric Sensors

  • Geometric sensors form the largest input-modality category, while learning methods can use monocular RGB images to generate reasonable motion without precise geometry.Geometric inputs include LiDAR, depth sensors, and sonars; RGB cameras offer low cost and semantic information, enabling behaviors such as lane keeping or turning away from a tree.Source: §5.2.1 · p. 23, §5.2.2 · p. 23

5.2.3 RGB + Geometry

  • RGB-plus-geometry methods use geometric sensing for obstacle avoidance and cameras for semantic information, while exteroceptive methods typically assume perfect perception in controlled settings.Together, the taxonomies highlight visual inputs as a departure from classical navigation’s primarily geometric framing.Source: §5.2.3 · p. 23, §5.2.4 · p. 23

6 ANALYSIS AND FUTURE DIRECTIONS

The analysis recaps the surveyed learning approaches by scope and navigation goals, then examines how learning scope relates to capabilities beyond classical navigation. It finds that many approaches were tested in simple settings without outperforming or even being compared with classical methods.

6.1.1 Recap

  • 43 of 74 surveyed papers used end-to-end approaches, including 15 that could not navigate to user-defined goals; the remaining 31 could navigate to those goals.Among the 31, 13 applied learning to a navigation subsystem (one global planner and 12 local planners), while 18 applied it to classical-system components (14 representations and four parameters).Source: §6.1.1 · pp. 23–24
  • 38/46 learning-based approaches addressing the classical navigation problem were tested in relatively simple environments and did not outperform—or were not compared with—classical approaches; 8/46 showed improvement.Source: §6.1.1 · p. 24
  • Of the 28/74 approaches extending beyond classical navigation, 24 addressed social navigation and four focused on terrain-based navigation.Source: §6.1.1 · p. 24

6.1.2 Cross-Dimensional Analysis

  • Among social-navigation works, 11 were end-to-end, 11 learned cost-function representations at component level, and only two were subsystem-level: [107] for global planning and for local planning.Source: §6.1.2 · p. 24, Figure 8 · p. 25
  • Terrain-based navigation had relatively few approaches, distributed across end-to-end, subsystem, and component categories.Source: §6.1.2 · p. 24, Figure 8 · p. 25

Social Navigation · COMPARISON

The comparison associates narrower learning scope with better outcomes, while the recommendations favor hybrid subsystem-level learning and call for stronger evaluation and new navigation capabilities.

COMPARISON

  • Narrower learning scope correlates with better results: no end-to-end approach improved on classical systems, while subsystem approaches span categories and three works showed gains.Component-focused methods mostly fall in the Improvement category, with novelty detection for collision prediction by Richter and Roy as the exception.Source: COMPARISON · p. 26

6.2 Recommendations and Future Directions

  • The authors recommend machine learning at the subsystem or component level, retaining classical components in hybrid systems for hard safety constraints.The recommendation follows the observed association between more limited learning scope and better overall system performance.Source: §6.2 · p. 26
  • Only 12 of 74 surveyed papers describe safety assurance, and each uses a classical-system component; learning methods typically lack explicit guarantees unlike bounded-uncertainty MPC.The authors identify this safety gap as a likely reason learning methods are rare in mission-critical systems.Source: §6.2 · p. 26
  • None of the 43 end-to-end approaches attempted explainability, while the other 31 surveyed approaches only partially maintained it through classical subsystems or components.Classical methods can make pipeline errors traceable and understandable, supporting diagnosis and improvement.Source: §6.2 · p. 26
  • Only six papers [104] studied machine-learning adaptation of classical systems; end-to-end research should now target performance beyond classical state of the art.The authors cite parameter tuning as one hybrid-system opportunity and note that end-to-end methods have repeatedly duplicated, but not convincingly advanced, classical navigation performance.Source: §6.2 · pp. 26–27
  • Evaluations should compare learning methods with classical baselines on real-world navigation metrics and disclose training and tuning costs, including the 73 of 74 approaches requiring manual hyperparameter search.One remaining approach automatically found parameters in 12 days; the authors also note data-collection and offline-training costs, often spanning days or weeks on high-end hardware.Source: §6.2 · p. 27
  • Further work should target reactive local planning: only four papers investigated local navigation under challenging constraints.The authors argue that local planning suits learning because it uses local sensor data for immediate outputs over limited spatial windows.Source: §6.2 · p. 28
  • Researchers should explore navigation behaviors beyond classical metric planning and develop systems that continually improve from deployment experience, a capability explicitly pursued by only one paper.The survey notes emerging social and terrain-aware navigation, dynamic parameter adjustment by APPL [104], and the potential for learning from deployment successes and failures.Source: §6.2 · p. 28

7 CONCLUSIONS

The article surveys machine-learning approaches to mobile robot motion planning and control through four perspectives linked to classical navigation, and identifies opportunities to combine learning-based and classical paradigms.

7 CONCLUSIONS

  • The survey organizes and analyzes prior work by its scope within navigation systems, comparison with classical methods, navigational tasks, and input modalities.These perspectives correspond to Sections 3, 4, 5.1, and 5.2.Source: §7 · p. 28
  • The article provides recommendations for the community and identifies promising future research directions in machine learning for autonomous navigation.Source: §7 · p. 28
  • Combining classical and learning-based approaches is identified as an opportunity to advance navigation toward faster, smoother, safer, and more reliable operation in more constrained, diverse, and challenging environments.Source: §7 · p. 29

Related papers

Loading 2011.13112v2…