PaperlayerView exact revision on arXiv

Source-linked AI summary

End-to-End Learning vs. Modular Architectures: Comparative Insights into Autonomous Driving Systems

Kartik B. Kapse

arXiv:2610.01746v1cs.ROcs.CVcs.LG

TL;DR

Direct comparisons of end-to-end and modular autonomous-driving architectures across key dimensions remain limited, complicating selection for specific applications. This paper presents a four-dimensional selection framework and finds that it correctly recommends the deployed architecture, exactly or partially, for 7/10 published systems.

  • Problem

    Comprehensive comparisons of end-to-end and modular architectures across scalability, robustness, and interpretability are limited, making application-specific framework selection challenging.

    Source: BASED MODULE · p. 4

  • Method

    The paper introduces a decision-support framework that uses twelve binary criteria across four dimensions to recommend an autonomous-driving architecture for a deployment context.

    Source: BASED MODULE · p. 4

  • Results

    7/10 published systems received an exact or partial framework recommendation matching their actual deployed architecture.

    Source: NVIDIA · p. 20

  • Takeaways & Limitations

    Architecture selection can be structured around application-specific safety, environment, resource, and deployment requirements.

    Source: BASED MODULE · p. 4, A. Comparative Overview · p. 16

    The framework is best suited to deployment-stage architecture decisions and is not intended to retrospectively classify early research prototypes.

    Source: NVIDIA · p. 20

Abstract

from arXiv · show

Autonomous driving systems have become a central focus of intelligent transportation research, with End-to-End Learning and Modular Architectures offering two prominent design paradigms for their implementation. E2E Learning uses deep learning algorithms to map raw sensory inputs directly to driving actuators, providing a streamlined and adaptable solution. while Modular Architectures employ a pipeline-based approach, dividing the system into distinct subsystems for perception, cognition, planning, and control. This paper presents a comprehensive comparative analysis of these paradigms, focusing on their strengths, limitations, and trade-offs to provide insights into their suitability for various autonomous driving applications. The study evaluates key factors such as interpretability, scalability, robustness, and real-world applicability. While End-to-End Learning emphasizes simplicity and adaptability in dynamic environments, it lacks transparency and is highly dependent on large datasets. Conversely, Modular Architectures offer superior interpretability and task-specific optimization, but face challenges related to integration complexity and scalability. To address these limitations, hybrid approaches that combine the strengths of both paradigms have emerged, offering a promising direction for overcoming these challenges. Beyond this comparative synthesis, following work proposes a Four-Dimensional Architecture Selection Framework, comprising twelve binary criteria across safety, operating environment, data/computational resources, and deployment context, and validate it against ten published autonomous driving systems, correctly recommending 7/10 deployed architectures. This work synthesizes existing literature to highlight key trade-offs between the paradigms and identifies hybrid architectures as a promising direction for future research.

I. INTRODUCTION

The paper compares modular, end-to-end, and hybrid autonomous-driving architectures across key design factors, and introduces a four-dimensional framework for selecting architectures by deployment context. Its retrospective validation reports correct or partial recommendations in seven of ten cases.

I. INTRODUCTION

  • Autonomous driving research is motivated by safety and traffic-flow benefits; NHTSA attributes approximately 94% of traffic accidents to human driver error.The introduction also emphasizes surveys’ role in synthesizing progress and anticipating developments.Source: §I · p. 1

BASED MODULE

  • Modular architectures support independent development and interpretability but face integration complexity, error propagation, and limited adaptation, while end-to-end systems are simpler but opaque and data-dependent.The paper evaluates the paradigms across interpretability, scalability, robustness, and real-world applicability.Source: BASED MODULE · pp. 2, 4
  • Imitation Learning trains models to replicate human actions but cannot cover every driving scenario, whereas Reinforcement Learning learns through environmental interaction without human driving data but is less data-efficient.A proposed hybrid training strategy first uses labelled data for imitation learning, then refines the policy through online reinforcement learning.Source: BASED MODULE · p. 3
  • The paper identifies a lack of comprehensive comparisons and standardized benchmarks, and highlights hybrid designs and cross-paradigm benchmarking, verifiable safety properties, and sim-to-real transfer as research directions.Hybrid approaches may balance modular interpretability with end-to-end adaptability.Source: BASED MODULE · pp. 4–5
  • The proposed selection framework maps twelve binary criteria across four dimensions to Modular, End-to-End Imitation Learning, End-to-End Reinforcement Learning, or Hybrid recommendations.The dimensions cover safety and regulatory compliance, operating environment, data and computational resources, and deployment and operational context.Source: BASED MODULE · pp. 4–5
  • Correct or partial architecture recommendations were demonstrated for 7 of 10 autonomous-driving systems from 1989–2023.The framework was compared retrospectively with each system’s deployed design.Source: BASED MODULE · p. 5
  • The review uses narrative synthesis of literature published from 1989 to 2025, drawing on IEEE Xplore, arXiv, Google Scholar, and MDPI rather than a formal systematic-review protocol.Sources were prioritized by citation count and venue, with peer-reviewed work favored alongside selected high-citation preprints.Source: BASED MODULE · p. 5

II. Autonomous Driving Paradigms · A. Autonomous Driving Paradigms Overview: · B. Modular Architecture

Autonomous driving systems use either end-to-end learning, which maps sensor inputs directly to vehicle commands, or modular architectures that process driving tasks through distinct stages. Modular systems offer interpretability and specialized optimization but face robustness and complexity challenges, while hybrid models seek to combine modular deliberation with end-to-end adaptability.

A. Autonomous Driving Paradigms Overview:

  • End-to-end learning maps raw sensor inputs directly to vehicle controls, while modular architectures divide driving into a sequence of processing and decision-making stages.The modular stages include sensing, perception, prediction/decision-making, planning, and control.Source: A. Autonomous Driving Paradigms Overview · p. 5
  • Modular architectures decompose driving into specialized stages, supporting task-specific optimization, interpretability, and maintainability.The overview identifies these as benefits of the modular breakdown.Source: Figure 3 · p. 6, A. Autonomous Driving Paradigms Overview · p. 6
  • Hybrid models combine modular interpretability and deliberate decision-making with end-to-end adaptive learning to unite reactive and deliberative behavior across driving scenarios.The passage notes modular systems’ particular effectiveness with map-based data and end-to-end models’ strength in processing complex sensory inputs.Source: A. Autonomous Driving Paradigms Overview · p. 6, Figure 3 · p. 6

B. Modular Architecture

  • Modular systems use well-defined intermediate representations and interdependencies, making errors easier to trace and allowing algorithms to be replaced without extensive changes elsewhere.They can also leverage methods and knowledge from computer vision and robotics [18].Source: B. Modular Architecture · p. 6
  • Modular systems can generalize poorly across diverse scenarios, overlook downstream-relevant context through engineered representations, and incur computational overhead from multiple models [18].For example, object detection may output only a bounding box and class label, omitting scene context that could help decision-making.Source: B. Modular Architecture · p. 6
  • Supervised-data dependence in perception models creates a human-information bottleneck, and errors in one module can propagate and compound downstream.Source: B. Modular Architecture · p. 7

a) Sensing … 2. Object Tracking

ADS combines complementary sensor data through fusion to support reliable environmental understanding, while perception processes that data for detection, tracking, and other driving tasks [20] [18].

a) Sensing

  • Sensor fusion combines cameras, LiDARs, radars, GPS, and IMUs to offset individual sensor weaknesses and provide a comprehensive view for real-time driving decisions.Cameras provide color and texture, LiDARs map depth, radars measure velocity and range, GPS positions the vehicle, and IMUs measure motion.Source: a) Sensing · p. 7

b) Perception

  • Perception processes sensor data into an environmental representation, with core tasks including object detection, tracking, and lane detection [20] [18].Source: b) Perception · p. 7

1. Object detection

  • Object detection is safety-critical because a missed vehicle can propagate errors to planning and control, potentially causing collisions.Source: §1 · p. 8
  • Two-stage detectors use a Region Proposal Network to generate regions of interest before classification and bounding-box regression [21].Source: §1 · p. 8
  • One-stage detectors such as YOLO and SSD directly regress densely sampled locations, prioritizing speed over accuracy.Source: §1 · p. 8
  • Object detection faces an accuracy-latency trade-off: more complex models tend to improve accuracy while increasing computational cost and latency.Source: §1 · p. 8

2. Object Tracking

  • Object tracking maintains consistent object identifiers across video frames to generate trajectories useful for trajectory prediction in ADS.Source: §2 · p. 8
  • SORT uses Kalman filtering and the Hungarian algorithm to maintain boxes and IDs, while deep trackers offer state-of-the-art performance at higher latency.Despite lower accuracy, lightweight SORT remains relevant for real-time applications because of its minimal computational overhead [18].Source: §2 · p. 8

3. Lane detection · 4. Performance Considerations and Challenges · c) Prediction

The sections describe lane-detection methods and perception trade-offs, then explain how prediction and planning use environmental context to support driving decisions. They emphasize computational demands, latency, and the propagation of perception errors as challenges for safe autonomous driving.

3. Lane detection

  • Lane detection faces variation in road topology, marking styles, and lane semantics; traditional geometric methods contrast with more robust but computationally demanding deep learning approaches.Traditional methods use color, texture, gradients, or energy minimization; deep learning approaches include segmentation models and CNN-RNN hybrids.Source: §3 · p. 8

4. Performance Considerations and Challenges

  • At 40 km/h in urban scenarios, control signals must be ready in less than 90 milliseconds, making the accuracy–latency trade-off critical for real-time decisions.Complex deep learning models can be highly accurate but introduce latency; traditional methods are less accurate but faster.Source: §4 · p. 8
  • Perception errors can propagate through the driving pipeline into suboptimal planning and control, potentially creating safety risks [31].Reliable operation in dynamic, time-sensitive environments also requires balancing detection speed and accuracy.Source: §4 · p. 8

c) Prediction

  • Prediction forecasts nearby vehicles’ and pedestrians’ future positions using tracked histories and scene context, while constant-velocity linear regression is less predictive but more computationally efficient than deep learning.Scene context can include LiDAR-derived 3D point clouds and bird’s-eye views; tracking provides past states for estimating velocity and lane occupancy.Source: c) Prediction · p. 9
  • Prediction becomes more computationally complex when multiple dynamic agents require simultaneous trajectory monitoring and inference.Predictions support proactive actions such as lane transitions and intersection navigation.Source: c) Prediction · p. 9
  • Planning uses predicted trajectories plus perception-derived localization and map data to generate waypoints through path, behavior, and motion stages; control translates decisions and trajectories into actuator commands.Control supports route adherence, speed modulation, safety, stability, and passenger comfort.Source: c) Prediction · pp. 9–10
  • Path planning seeks a minimum-cost, safe route that accounts for distance, comfort, safety, energy, collision risk, obstacles, and traffic rules, but rule-based modular methods may adapt poorly to real-world variability.Source: c) Prediction · p. 9
  • Behavior planning selects high-level actions using MDPs or POMDPs under uncertainty, while motion planning converts the chosen action into a feasible trajectory using graph or incremental search [35].Behavior actions include lane keeping and preparatory lane changes; motion trajectories aim to be safe, comfortable, and dynamically feasible.Source: c) Prediction · pp. 9–10

C. End-to-End Learning · a) Imitation Learning

End-to-End systems map raw sensory inputs directly to control commands, simplifying architecture but making behavioral outputs harder to interpret. This section examines imitation learning, which learns from expert demonstrations but can inherit their limitations and struggle beyond demonstrated behavior.

C. End-to-End Learning

  • End-to-End driving maps raw sensory inputs directly to control commands and uses fewer, simpler components than modular systems, but makes failures harder to explain.Unlike modular systems, where failures can be localized to a constituent module, End-to-End behavior can be difficult to diagnose when anomalies occur.Source: C. End-to-End Learning · p. 10

a) Imitation Learning

  • Imitation learning reproduces expert or reference-model actions, leveraging human expertise and often producing safer, more interpretable behavior than reinforcement learning, but inherits demonstration errors and cannot readily exceed their repertoire.End-to-End methods are divided into imitation and reinforcement learning.Source: a) Imitation Learning · p. 10, C. End-to-End Learning · p. 10
  • Imitation learning trains a policy to map observations to expert actions through supervised learning, then evaluates reproduction on unseen data or task completion under varied conditions.Evaluation may compare the agent’s actions directly with the expert’s or assess task performance.Source: a) Imitation Learning · p. 11
  • Imitation learning comprises behavior cloning and inverse reinforcement learning; its formulation uses state-action trajectories to learn a policy that mimics the expert policy.The dataset contains sequences of states and expert actions.Source: a) Imitation Learning · p. 11
  • Behavior cloning frames imitation as supervised policy alignment; ALVINN used a fully connected network for lane following, while modern methods use deep networks [12].Because behavior cloning imitates only training-set actions, covariate shift can compound errors in new states; data augmentation and regularization add noise to improve robustness and generalization.Source: a) Imitation Learning · p. 11
  • Inverse reinforcement learning infers a latent reward from expert trajectories and state features, then uses methods such as Q-learning or policy gradients to derive a policy maximizing that reward.It is useful when manually specifying a reward is intractable, and formulates reward recovery as constrained optimization.Source: a) Imitation Learning · pp. 11–12
  • IRL faces ambiguous reward inference from noisy, suboptimal human demonstrations, high data and computational demands in high-dimensional driving states, and poor generalization across different environments.The passage contrasts highway driving with urban intersections as an example of limited cross-domain applicability.Source: a) Imitation Learning · p. 12

b) Reinforcement Learning

Reinforcement learning (RL) develops driving decisions through iterative interaction with an environment, maximizing cumulative reward while balancing exploration and exploitation. The section outlines value-based, policy-based, and actor-critic methods, their limitations, and their applications to autonomous driving.

Reinforcement Learning

  • RL faces challenges in efficient exploration, long-term reward handling, and safe real-world learning, while policy-based methods can also suffer unstable gradient updates.Exploration strategies include ϵ-greedy and SoftMax, with more exploration typically used early in training; policy methods use stochastic policies and tend to offer smoother updates.Source: b) Reinforcement Learning · pp. 12–13

Value-Based Reinforcement Learning

  • Value-based RL estimates state or action values; Q-learning updates Q-values, and DQN applies neural networks to high-dimensional states and supports highway lane-change decisions.DQN is suited to discrete actions such as staying in lane or changing left or right.Source: b) Reinforcement Learning · pp. 12–13

Policy-Based Reinforcement Learning

  • Policy-based RL directly optimizes a parameterized state-to-action policy; TRPO limits updates using KL divergence, while PPO clips its objective to avoid excessive policy deviations.PPO is widely used in CARLA-based training because stable updates suit the long episodes typical of driving tasks.Source: b) Reinforcement Learning · p. 13

Actor-Critic Reinforcement Learning

  • Actor-critic methods pair a policy-learning actor with a value-estimating critic, using an advantage function to guide updates; DPG [49] learns deterministic policies with fewer samples in large action spaces.Actor-critic methods suit continuous driving controls such as steering angle and throttle because they naturally handle continuous action spaces.Source: b) Reinforcement Learning · p. 13

c) Teacher-Student Paradigm

The teacher-student paradigm trains a privileged Teacher to guide a sensor-only Student, but faces challenges from noisy inputs, behavior-cloning causal confusion, and inefficient knowledge distillation.

c) Teacher-Student Paradigm

  • In, a privileged agent first learns from environmental state and then supervises a vision-only sensorimotor agent through imitation.The two-stage process lets the Student develop sensorimotor capabilities while benefiting from the Teacher’s guidance.Source: Figure 5 · p. 14, c) Teacher-Student Paradigm · p. 14
  • [53] identifies the Student’s independent planning head as difficult to learn from noisy, redundant sensory input and behavior-cloning causal confusion.The proposed approach exploits the Teacher’s stronger planning while letting the Student focus its representational capacity on perception.Source: c) Teacher-Student Paradigm · p. 14
  • Knowledge-distillation inefficiencies remain a challenge despite advances in Teacher robustness and knowledge transfer.Source: c) Teacher-Student Paradigm · p. 14

d) End to end Autonomous driving with the auxiliary task · D. Hybrid Architecture

Auxiliary tasks add structure and interpretability to end-to-end driving, while the hybrid design assigns perception and control to modular components and data-driven path planning to end-to-end methods. The hybrid system performed well in both mapless and map-based tracks at the 2023 CARLA Autonomous Driving Challenge.

d) End to end Autonomous driving with the auxiliary task

  • For end-to-end learning, one approach trains perception and planning independently in two sequential stages to reduce computational and representational complexity [53].Source: d) End to end Autonomous driving with… · p. 14
  • Uber ATG’s interpretable end-to-end planner selects the lowest-cost candidate trajectory using a cost volume derived from LiDAR and map inputs.The backbone also predicts bounding boxes, and a sampling module generates candidate trajectories whose costs are evaluated against the cost volume.Source: d) End to end Autonomous driving with… · pp. 14–15
  • The planner uses detection, motion-forecasting, and human-trajectory supervision, with a max-margin loss separating good trajectories from sampled negative candidates [9].Perceptual loss combines classification, regression, and IoU losses on bounding boxes, while the cost volume receives no direct supervision [9].Source: Figure 6 · p. 15, d) End to end Autonomous driving with… · p. 15
  • Another end-to-end approach uses segmentation masks, monocular depth, and optical flow as intermediate imitation-learning signals to guide control and improve resilience and interpretability.Source: d) End to end Autonomous driving with… · p. 15

D. Hybrid Architecture

  • The hybrid approach addresses modular systems’ complexity and error propagation alongside end-to-end systems’ limited transparency and need for large-scale data.Modular systems are suited to well-defined situations requiring interpretable components, whereas end-to-end systems are simpler and adaptable to dynamic scenarios.Source: D. Hybrid Architecture · p. 15
  • The hybrid architecture combines modular interpretability with end-to-end adaptability, assigning perception and control to modular components and data-driven path planning to end-to-end methods.It uses lidar, stereo cameras, and other sensors to detect obstacles and traffic lights, then fuses sensor data with BEVFusion for CNN-based waypoint planning.Source: D. Hybrid Architecture · pp. 15–16
  • The hybrid system performed well in both SENSOR (mapless) and MAP tracks at the 2023 CARLA Autonomous Driving Challenge.Its modular design enabled component-level debugging, while data-driven components made real-time decisions under dynamic conditions.Source: D. Hybrid Architecture · p. 16

III. DISCUSSION · A. Comparative Overview

The paper proposes a context-specific framework for choosing among autonomous-driving architectures and validates it against ten published systems, finding exact or partial matches for seven. Its criteria link architecture suitability to safety, environment, resources, and deployment context rather than assuming one paradigm is universally best.

A. Comparative Overview

  • Environment-modeling strategy and system architecture are orthogonal axes: model-based methods use explicit environment representations, while modular systems divide driving into components and E2E systems unify perception and control.The axes can interact; for example, a modular system may use model-free methods in planning.Source: A. Comparative Overview · p. 16
  • The framework uses twelve yes-or-no criteria across four dimensions, recommending the architecture with the most accumulated signals for the application’s requirements.A yes counts toward every architecture named by a criterion; broad E2E signals count only toward Hybrid, while subtype-specific criteria distinguish E2E imitation learning from reinforcement learning.Source: A. Comparative Overview · p. 16, Table 2 · p. 19, NVIDIA · p. 19
  • Safety criteria favor Modular or Hybrid when traceability, component-level failure isolation, or formal certification is required; structured environments favor Modular, while unpredictable, edge-case-rich settings favor E2E or Hybrid.The safety dimension includes legal decision-traceability requirements, isolating failures to a component, and mandated safety-integrity certification.Source: A. Comparative Overview · p. 16
  • Scarce data, simulator, or onboard-compute resources favor Modular; production systems needing independent component updates benefit from modularity, while Hybrid suits teams with classical-control and deep-learning expertise.E2E imitation learning requires large labelled datasets, reinforcement learning requires a capable simulator, and both are computationally expensive at inference.Source: A. Comparative Overview · pp. 16–17

NVIDIA

  • No architecture dominates across criteria: Modular suits safety-critical, resource-constrained deployments; E2E suits data-rich, unstructured settings; RL suits simulation-heavy research; and Hybrid balances interpretability and adaptability.The framework’s mapping of ten published systems supports the paper’s conclusion that application context, rather than algorithmic superiority alone, drives architecture choice.Source: NVIDIA · p. 19, Table 3 · p. 19, Table 4 · p. 19
  • 7/10 published systems received exact or partial deployed-architecture matches; four were exact—Boss, Apollo, Uber ATG NMP, and CaRINA Agent—and TransFuser, UniAD, and LAV were partial.The three partial matches were End-to-End Imitation Learning systems with auxiliary modular-style supervision, making them architecturally close to Hybrid under the paper’s definition.Source: Table 4 · p. 19, NVIDIA · p. 20
  • ALVINN and PilotNet mismatches reflect early research prototypes outside the framework’s deployment-stage scope, while Roach’s Hybrid plurality (2 to 1 over E2E-RL) mismatched its pure E2E-RL deployment.The authors treat early-prototype mismatches as a scope boundary; Roach lacked auxiliary modular-style supervision, so the partial-match rationale for TransFuser, UniAD, and LAV does not apply.Source: NVIDIA · p. 20

a) Environment Modeling Paradigms: Model-Based vs. Model-Free · 1. Model-Based Approaches

Model-based approaches use vehicle and environment models to predict motion, optimize trajectories, and support stable control. Examples include MPC, graph-based path planning, and physics-based vehicle dynamics models.

1. Model-Based Approaches

  • MPC predicts future vehicle states from a dynamic vehicle and environment model, then optimizes a trajectory over a finite horizon for safe, efficient path planning.Source: §1 · p. 20
  • Google’s Self-Driving Car project used an MPC framework to handle complex vehicle dynamics and optimize control signals for smooth driving.Source: §1 · p. 20
  • D* and A* compute optimal paths on road-network graphs using predefined models of the environment and vehicle motion.Source: §1 · pp. 20–21
  • Tesla’s Autopilot uses model-based planning to anticipate nearby vehicle movements and plan optimal trajectories.Source: §1 · p. 21
  • Physics-based vehicle dynamics models use kinematic and dynamic models to predict motion and maintain stability during maneuvers.Source: §1 · p. 21
  • Examples include modeling tire-road interactions or lateral and longitudinal dynamics, as used by Roborace autonomous race cars for high-speed control.Source: §1 · p. 21

2. Model-Free Approaches … 2. End-to-End Learning

The section surveys model-free driving methods and contrasts modular and end-to-end design tradeoffs: end-to-end systems streamline sensor-to-control learning but face opacity, data, and safety concerns, while modular systems aid interpretability but encounter integration and generalization limits.

2. Model-Free Approaches

  • Deep Q-learning trains agents to select actions by optimizing rewards through environmental interaction without a predefined model; DQN examples include highway lane changes and obstacle avoidance.Source: §2 · p. 21
  • PPO teaches driving behaviors in simulation, with CARLA experiments covering intersections and pedestrian crossings, while PilotNet maps camera input to steering by imitating expert demonstrations [12].Source: §2 · p. 21

1. Modular Architecture

  • Modular systems aid debugging through interpretable components, allow independent updates, and use MPC for robust performance in structured driving environments.Source: §1 · p. 21
  • Modular systems can suffer integration misalignment and cascading errors, while reliance on handcrafted features limits generalization in dynamic, unstructured environments.An incorrect perception output can affect downstream planning and control.Source: §1 · p. 21

2. End-to-End Learning

  • End-to-end systems map raw sensor inputs directly to control outputs, reducing architectural complexity; with diverse training data, they can generalize across environments and improve through data-driven optimization.Source: §2 · pp. 21–22
  • End-to-end models are difficult to interpret, making their decisions harder to debug and complicating troubleshooting and regulatory compliance.Source: §2 · p. 22
  • End-to-end training requires extensive, diverse labeled data, including rare edge cases, and insufficient diversity can cause overfitting and poor performance in unfamiliar situations.Such datasets can be difficult and expensive to collect.Source: §2 · p. 22
  • Without explicit intermediate checks or redundancy, end-to-end systems face challenges ensuring robustness and reliability in critical situations, risking failures in high-stakes scenarios.These concerns arise from the lack of built-in safeguards.Source: §2 · p. 22

B. Trade-offs: … IV. CONCLUSION:

The section contrasts modular systems’ transparency and structured-environment strengths with end-to-end systems’ adaptability and data-driven scalability, while noting their respective limitations. It presents hybrid designs and a Four-Dimensional Architecture Selection Framework as contributions toward addressing these trade-offs.

a) Transparency vs. Flexibility

  • Modular components are transparent for accountability and failure analysis, while end-to-end systems adapt better to complex, unstructured situations but obscure why high-performing CNNs fail.The passage characterizes this as a trade-off between performance and durability.Source: a) Transparency vs. Flexibility · p. 22

b) Scalability

  • End-to-end performance improves with larger simulation datasets, but real-world scaling faces data-collection costs and domain gaps; modular systems can avoid retraining in predictable, structured settings.Source: b) Scalability · p. 22

c) Error Diagnosis

  • Modular systems enable quicker debugging by isolating errors in components such as perception or planning, whereas interconnected end-to-end systems make fault localization and fixes more complex.Source: c) Error Diagnosis · p. 22

C. Future Trends:

  • Hybrid designs use modular components for interpretability and precision-critical tasks, and end-to-end learning for data-benefiting tasks, supporting adaptability, interpretability, and safer, more robust generalization.Examples include modular perception and end-to-end planning, or learned perception and motion forecasting paired with modular trajectory optimization.Source: C. Future Trends · pp. 22–23

IV. CONCLUSION:

  • The framework correctly identified deployed architectures in 7/10 systems—four exact matches and three partial matches—and is intended for deployment-stage decisions, not retrospective classification of early prototypes.It was validated against ten systems deployed between 1989 and 2023.Source: §IV · p. 23
  • Modular systems offer transparency, adaptability, and reliability in structured environments but face integration, error-propagation, and generalization challenges in dynamic settings; end-to-end systems are data-hungry and hard to interpret, with overfitting and safety concerns.Source: §IV · p. 23
  • Future work should establish fair cross-benchmark evaluation, develop hybrids with formally verifiable safety, and reduce the sim-to-real gap for end-to-end policies through domain randomization and adaptive transfer.Source: §IV · p. 23

Related papers

Loading 2610.01746v1…