Source-linked AI summary
Motion Planning for Autonomous Driving: The State of the Art and Future Perspectives
Siyu Teng, Xuemin Hu, Peng Deng, Bai Li, Yuchen Li, Dongsheng Yang, Yunfeng Ai, Lingxi Li, Zhe Xuanyuan, Fenghua Zhu, Long Chen
TL;DR
Autonomous-driving motion planning still faces reliability, safety, data, interpretability, and generalization challenges that limit real-world deployment. This survey reviews pipeline and end-to-end planning, including their operations, learning methods, platforms, comparisons, challenges, and future perspectives. It identifies pipeline and end-to-end planning as the two broad categories while covering imitation, reinforcement, and parallel learning within end-to-end methods.
Problem
Intelligent vehicles remain largely limited to small-scale trials, while safe, robust, and adaptable planning is needed for complex scenarios and real-world deployment.
Method
The paper surveys pipeline and end-to-end motion-planning methods, examining pipeline operations, end-to-end training approaches, validation scenarios, experimental platforms, comparisons, challenges, and future perspectives.
Results
The survey classifies autonomous-driving planning into pipeline and end-to-end categories and reviews imitation learning, reinforcement learning, and parallel learning for end-to-end planning.
Takeaways & Limitations
The comparison of pipeline and end-to-end methods supports system-level design choices by clarifying their strengths, limitations, and development strategies.
Takeaways & Limitations
End-to-end planners remain scarcely implemented in the real world because of challenges involving data, interpretability, generalization, and policies.
Abstract
from arXiv · showhide
Intelligent vehicles (IVs) have gained worldwide attention due to their increased convenience, safety advantages, and potential commercial value. Despite predictions of commercial deployment by 2025, implementation remains limited to small-scale validation, with precise tracking controllers and motion planners being essential prerequisites for IVs. This paper reviews state-of-the-art motion planning methods for IVs, including pipeline planning and end-to-end planning methods. The study examines the selection, expansion, and optimization operations in a pipeline method, while it investigates training approaches and validation scenarios for driving tasks in end-to-end methods. Experimental platforms are reviewed to assist readers in choosing suitable training and validation strategies. A side-by-side comparison of the methods is provided to highlight their strengths and limitations, aiding system-level design choices. Current challenges and future perspectives are also discussed in this survey.
I. INTRODUCTION
Autonomous vehicles remain largely limited to trial programs despite their potential benefits, making safe and robust motion planning essential. The introduction contrasts modular pipeline frameworks with end-to-end learning approaches and outlines their respective strengths and limitations.
- Intelligent vehicles attract broad interest because they may improve transportation, reduce road accidents, alleviate congestion, and improve urban mobility.
- Despite advances in artificial intelligence, hardware, and sensing, intelligent vehicles remain primarily confined to limited trials because reliability and safety remain concerns.Sensor limitations also hinder detection and response in complex scenarios.
- Pipeline planning uses interconnected modules and integrates with perception, localization, and control to support autonomous driving.It includes global route planning and local behavior or trajectory planning.
- Pipeline frameworks are interpretable and support fault localization, but require substantial computational resources and may not optimize every driving scenario.
- End-to-end planning treats driving as one learnable task, potentially improving interaction efficiency, generalization, robustness, and real-time capability.Its main limitation is that model errors and control decisions are more difficult to trace and explain.
C. Paper Structure
The paper organizes autonomous-driving motion planning into pipeline and end-to-end categories, then reviews their methods and supporting development platforms. Its survey emphasizes industry-relevant pipeline classification and newer end-to-end learning categories.
- Autonomous-driving planning methods are broadly classified into pipeline and end-to-end approaches.
- The survey addresses a prior lack of comprehensive review covering both pipeline and end-to-end motion-planning methods.
- Pipeline methods are classified by expansion and optimization mechanisms, including state grid identification, primitive generation, and other approaches.
- End-to-end planning coverage includes imitation learning, reinforcement learning, and the introduced parallel-planning category.The parallel-planning category uses virtual-real interaction confusion learning for end-to-end planning.
- Pipeline planning is reviewed through global route planning and local behavior or trajectory planning, while route planners search road-network graphs for minimal-cost paths.
B. Local Behavior/Trajectory Planning
Local behavior and trajectory planning combines route-following decisions with trajectory generation under vehicle dynamics, constraints, and moving obstacles. Because the underlying optimal-control problem generally lacks an analytical solution, planners construct trajectories through state-grid and primitive operations.
- Local behavior and trajectory planning jointly compute safe, comfortable, continuous local trajectories from identified global routes.Because the planning horizon is local, implementation generally uses a receding-horizon strategy.
- The local-planning optimal-control formulation minimizes a cost function subject to vehicle dynamics, state and control bounds, endpoint conditions, and workspace constraints.
- When the optimal-control problem lacks an analytical solution, trajectory construction uses state-grid identification and primitive generation, or combines both operations.
- State-grid identification can use search, selection, optimization, or potential minimization; search-based methods include A* and dynamic programming.
- Optimization-based state-grid methods discretize the optimal-control problem into mathematical programs whose solutions provide high-resolution state grids.Solvers may be gradient-based or non-gradient-based, and methods can be combined into coarse-to-fine planning strategies.
2) Primitive Generation:
Primitive generation supplies trajectory or path elements for local planning through analytical, simulation, interpolation, and optimization approaches. In end-to-end planning, learning methods instead map sensor information toward trajectories or controls, with imitation learning and reinforcement learning offering distinct training paradigms.
- Primitive Generation: Primitive generation commonly uses closed-form rules, simulation, interpolation, or optimization to construct trajectory or path primitives.Examples of closed-form methods include Dubins or Reeds-Shepp curves, polynomials, and theoretical optimal control.
- Primitive Generation: State-grid identification and primitive generation are fundamental trajectory-construction operations that may be arranged iteratively, offline-before-online, or sequentially.
- End-to-End Planning: End-to-end planning directly maps raw sensor data to trajectory points or control signals without an external perception-control gap or many human-designed heuristics.
- Imitation Learning: Imitation learning trains a policy from expert state-action trajectories, using state-action pairs to learn the relationship between states and actions.
- Imitation Learning: Behavioral cloning, direct policy learning, and inverse reinforcement learning are identified as three widely used imitation-learning approaches.
1) Behavioral Cloning:
Behavioral Cloning (BC) learns driving control policies by passively mapping expert state-action examples, but its coverage and generalization depend strongly on the demonstrations. Direct Policy Learning and DAgger-based extensions add online data collection or recovery guidance to address these limitations.
- BC trains a classifier or regressor to replicate policies from expert trajectories through passive observation of complete commands.
- Conditional imitation learning combines monocular images, ego-vehicle velocity, and high-level commands to produce lateral and longitudinal control.
- Additional route, location, or point-cloud inputs improve reported generalization and robustness by providing richer perception data.
- BC generalization is limited when expert data are biased or overfit, and recovery behavior is difficult in states absent from demonstrations.
- DPL evaluates the current policy to obtain more suitable training data, while DAgger aggregates experienced state-action pairs through online learning.
- DAgger-based methods reduce dataset dependency and improve learning efficiency, but they may suffer from inefficient querying, inaccurate collection, poor generalization, and inability to distinguish expert trajectory quality.
- DPL supports iterative policy improvement by reducing dataset volume and distribution requirements while eliminating incorrect policies.
3) Inverse Reinforcement Learning:
Inverse Reinforcement Learning (IRL) infers reward structure from expert trajectories before optimizing behavior, offering alternatives to direct state-action imitation. The reviewed variants include max-margin, Bayesian, and maximum-entropy methods, alongside interpretable and preference-based planning systems.
- IRL infers latent reasons connecting inputs and outputs, then optimizes behavior using an inferred reward function rather than only a state-action mapping.
- Max-margin methods: Max-margin IRL learns feature-weighted rewards that maximize the margin between an optimal policy and estimated sub-optimal policies.
- Interpretable and preference-based systems: An interpretable planning system generates trajectories, applies a safety filter, and scores candidates with DeepIRL; preference inference additionally supports user-selected actions.
- Bayesian methods: Bayesian IRL estimates posterior reward distributions from prior reward assumptions or optimized trajectories, with sampling methods addressing unseen scenarios.
- Maximum-entropy methods: Maximum-entropy IRL is suited to continuous spaces and can mitigate noise, imperfect behavior, and sub-optimal effects in expert trajectories.
- Maximum-entropy methods: GAIL uses a generative adversarial network to reconstruct expert-trajectory distributions and achieves human-comparable performance in specific scenarios.
- IRL methods retain long-tail corner-case problems, while expert-trajectory quality and distribution can constrain performance and dataset-based generalizability.
B. Reinforcement Learning
Reinforcement learning addresses the data demands and training uncertainty of imitation learning by learning policies through environment interaction. The review formulates driving RL with Markov decision processes and identifies value-based, policy-based, hierarchical, and multi-agent approaches.
- RL learns policies by interacting with an environment rather than imitating manually labeled expert data.
- An MDP formulates RL with state and action spaces, a reward function, transition probabilities, and a discount factor γ.
- At each time step, the agent selects an action from its policy, transitions to a new state, and receives rewards it seeks to maximize over time.
- The reviewed RL approaches include value-based and policy-based learning, with hierarchical and multi-agent extensions proposed for more complex driving scenarios.
1) Value-based Reinforcement Learning:
Value-based reinforcement learning estimates action values from expected rewards and has been applied to discrete driving decisions, while continuous-control and training-efficiency limitations motivate other approaches.
- Value-based methods estimate action values from expected rewards in each state and use them to select decisions.
- Q-Learning and DQN methods map image or state information to driving maneuvers, discrete actions, or trajectory-planning targets.
- The value-based approach is limited to discrete commands, whereas autonomous driving requires continuous, fine-grained control.
- DDPG achieved human-level lane-following performance from monocular images during a 250m real-vehicle road test.
- PPO was used for continuous motion planning and automated lane-change policies in congested roads and real highway scenarios.
- Combining reinforcement learning with imitation, curriculum learning, or human guidance is presented as a response to inefficient exploration and difficult training.
- Driving-task performance may require combining multiple methods and designing task-specific training methods.
3) Hierarchical Reinforcement Learning:
Hierarchical reinforcement learning decomposes autonomous-driving problems into subtasks with separate goals and policies, targeting the training difficulty of complex, high-dimensional scenarios.
- 3) Hierarchical Reinforcement Learning: HRL decomposes the total driving problem into hierarchically organized subtasks, each with its own goal and policy.
- 3) Hierarchical Reinforcement Learning: Two-level lane-changing methods assign high-level networks to decide whether to change lanes and low-level networks to execute commands.
- 3) Hierarchical Reinforcement Learning: An HRL approach for complex dynamic traffic uses separate high-level and low-level planning layers for autonomous decision-making and motion planning.
- 3) Hierarchical Reinforcement Learning: Cola-HRL combines a high-level planner, low-level controller, and continuous-lattice state representation.
- 3) Hierarchical Reinforcement Learning: Cola-HRL outperforms other state-of-the-art methods for high-quality decisions across various scenarios.
- 4) Multi-Agent Reinforcement Learning: Multi-agent reinforcement learning addresses interactions among traffic participants, including cooperative, competitive, mixed-traffic, and collision-avoidance settings.
- 4) Multi-Agent Reinforcement Learning: Shapley-value reward reallocation is used to motivate stable cooperation among intelligent vehicles.
- Deep reinforcement learning methods remain challenged by poor sample efficiency and limited interpretability.
C. Parallel Learning
Parallel learning combines artificial-system data, computational experiments, and real-world interaction to address data scarcity, inefficient learning, robustness, and emergency-planning demands in end-to-end driving.
- Planning methods face heuristic coupling, expert-trajectory requirements, and computational demands that impede widespread autonomous-driving implementation.
- CPSS links virtual and real systems through interaction and feedback, using real data to construct artificial systems and artificial systems to support real operations.
- Parallel system theory centers on the ACP combination of artificial societies, computational experiments, and parallel execution.
- Parallel learning is proposed as a framework for extending neural-network learning and addressing imitation-learning and reinforcement-learning challenges.
- Parallel driving bridges expert trajectories and control commands through neural-network learning for scenario-specific policies.
- Parallel planning generates expert trajectories from artificial traffic scenarios using pretrained knowledge from reality to address data insufficiency.
- Parallel planning supports rational emergency decisions without a heavy calculation burden.
- Parallel driving is described as alleviating data shortage, inefficient learning, and poor robustness in end-to-end planning models.
IV. EXPERIMENT PLATFORM
Experimental platforms support autonomous-driving training and validation because real-system testing carries potentially fatal safety risks and end-to-end models depend strongly on dataset characteristics.
- Real-system testing can involve potentially fatal safety risks, so autonomous-driving algorithms are often evaluated in artificial systems using datasets and simulations.
- End-to-end training requires large volumes of human-driving data, whose magnitude, abundance, and distribution affect safety, robustness, and generalization.
- KITTI supports object detection, sceneflow, depth estimation, tracking, and related perception tasks through broad task scaling.
- Comma.ai adds localization information and control signals, enabling localization and planning tasks.
- BDD100K and SODA10M provide urban scenarios across more than 31 cities under varied weather conditions with multiple label types.
- A2D2 provides 360° point-cloud perception using five LiDARs for full-scene autonomous-driving perception.
- Automine supplies 18 hours of videos and localization data from six open-pit mines with rough terrain, intense light, and dust.
- Datasets are described as crucial for training and validating intelligent-vehicle methods.
B. Simulation Platform
Simulation platforms provide low-cost, safer validation for autonomous-driving algorithms across traffic, urban, racing, and customizable scenarios. The reviewed platforms differ in supported agents, environments, sensor configurations, and development purposes.
- B. Simulation Platform: Simulation testing accelerates autonomous-driving validation because it offers lower cost and higher safety than real-world testing.It is used to validate algorithms while reducing the risks associated with real-world scenarios.
- B. Simulation Platform: SUMO supports large-scale transportation-algorithm validation through microscopic traffic simulation and broad data-format compatibility.Its interface is designed to handle diverse data formats for transportation simulations.
- B. Simulation Platform: Other platforms cover large-scale data interaction, corner-case generation, vehicle testing, varied environments, and single- or multi-vehicle validation.The reviewed examples include Drive Constellation, SVL Simulator, CarMaker, AriSim, Apollo, and Autoware.
- B. Simulation Platform: CARLA targets urban autonomous driving, while MetaDrive supports customizable road segments and generalizable reinforcement-learning research.CARLA supports development, training, and validation; MetaDrive is compositional and suited to varied complex scenarios.
- B. Simulation Platform: TORCS supports multi-vehicle interaction with more than 50 vehicle models, over 20 tracks, and simultaneous racing against 50 vehicles.Its ability to run many agents is useful for dense-vehicle scenarios and scheduling functions.
C. Physical Platform
Physical validation remains necessary because simulation cannot represent every corner case and simulated models often require real-world fine-tuning. Autonomous driving has reached semi-open-road validation, but complete commercial deployment still faces planning, safety, perception, dataset, and end-to-end limitations.
- C. Physical Platform: Simulation can address many scenarios and long-tail cases, but models typically require fine-tuning before real-world implementation.Simulation also cannot account for all corner cases, motivating professional semi-open validation sites.
- C. Physical Platform: Autonomous-driving systems have been validated on semi-open roads in multiple cities, while complete commercial deployment remains unrealized.The paper attributes the remaining gap to numerous obstacles and impending challenges.
- C. Physical Platform: Perception sensors remain vulnerable to environmental effects and partial perception, creating security challenges when potential hazards are missed.The paper identifies perception limitations as one of the central challenges for intelligent vehicles.
- C. Physical Platform: Virtual-environment training does not always transfer directly to reality, making the virtual–real data gap a priority for autonomous-driving research.The paper specifically identifies dataset transfer between simulators and real environments as an imperative research issue.
- C. Physical Platform: End-to-end planners remain scarcely implemented in the real world because of challenges involving data, interpretability, generalization, and policies.The paper presents future perspectives for addressing these barriers.
- C. Physical Platform: Interpretability, Sim2Real alignment, and reliability are identified as future directions for improving end-to-end planning and intelligent-vehicle deployment.The paper links black-box inference and simulation–reality differences to trust, safety, data diversity, and fairness concerns.