AUTO-DISCERN: Autonomous Driving Using Common Sense Reasoning

1 Introduction

• The proposed division uses machine learning for visual pattern matching and scene understanding, while automated commonsense reasoning makes driving decisions.The authors argue that over-reliance on machine learning for all aspects of autonomous driving has slowed progress.Source: §1 · p. 1
• The system automates commonsense reasoning with answer set programming and its goal-directed implementation, s(CASP).The authors say goal-directed reasoning matters because SAT-solver-based systems such as CLINGO face practical scalability and explainability issues for autonomous driving.Source: §1 · p. 1

2 Autonomous Vehicle Technology

• AUTO-DISCERN takes a scene description and sensor values as input and computes a driving decision for that instant using commonsense rules.The authors expect deep-learning image processing to provide the scene description; they say decisions are explainable and, in theory, never wrong if the rules and system inputs are correct.Source: §2 · pp. 1–2
• In this paper, the authors identify their main contribution as showing how automated commonsense reasoning can be harnessed to achieve an SAE Level 5 autonomous driving system.Source: §2 · p. 2
• Although autonomous vehicles promise transportation and mobility benefits, the paper notes that none had earned the SAE Level 5 designation.The passage also contrasts earlier projections of widespread adoption with the status described by the authors.Source: §2 · p. 2

3 Machine Learning-based AV Systems

• In the Figure 2 lane-merge example, radar detects a left-side barrier, but reflection confuses the visual component, which misses that the lane is ending.The authors use the example to illustrate an ML-based AV failure.Source: Figure 2 · p. 2, §3 · p. 2
• Machine-learning systems can predict steering, braking, and acceleration from driving data, while rare or unseen situations may fall outside what their training data captures.The paper describes the learned mapping as statistical and cites flashing emergency lights and erroneous scene interpretation as examples of difficulty.Source: §3 · p. 2
• Figure 3 illustrates traffic-sign misclassifications after partial modifications, including tape that changes a detected 35 speed limit to 85.The paper presents sign perturbations as another failure case for machine-learning models.Source: Figure 3 · p. 3, §3 · p. 3
• The authors argue that AUTO-DISCERN can safely handle situations where machine-learning-based systems fail, while companies have scaled back or restricted Level 5 ambitions.Examples given include moving to SAE Level 4 or limiting autonomous behavior to geo-fenced areas.Source: §3 · p. 3

Motivation

• The paper distinguishes visual scene processing and inference from vehicle control, noting that machines find control easier than visual inference.The human-driving capabilities it lists include deductive and abductive inference and distinguishing real objects from depicted ones.Source: Motivation · p. 3
• The proposed division uses machine learning for scene processing and automated commonsense reasoning to infer driving actions such as steering, accelerating, or braking.The paper argues that pattern-matching tasks suit machine learning, while deduction tasks suit automated reasoning.Source: Motivation · p. 3
• Deep-learning image processing is expected to label scene objects and bounding boxes as predicates that commonsense rules use to compute a driving decision.The paper identifies datasets such as Visual Genome and systems such as DenseCap as possible sources or tools for labeling images.Source: Motivation · p. 3

• Commonsense reasoning revises conclusions drawn from defaults when new information reveals an exception, as when a presumed driver is later identified as a robot-taxi.The paper also describes preferences as a way to resolve which default applies when multiple defaults are available.Source: §5 · pp. 3–4
• ASP can represent defaults, exceptions, preferences, and multiple worlds, while s(CASP) executes these representations.Multiple worlds help distinguish, for example, an actual car from a car shown on a billboard.Source: §5 · p. 4
• SAT-solver-based ASP implementations such as CLINGO require grounding and may incur exponential program growth, compute entire models, and lack conclusion justifications.These are the implementation problems contrasted with goal-directed ASP.Source: §5 · p. 4
• Goal-directed s(CASP) explores only knowledge needed to answer a query and provides a proof tree that justifies the answer.It works directly on predicate ASP programs without grounding them first.Source: §5 · p. 4

6 Commonsense Reasoning-based AV

• For each timestamp, scene facts and driving rules combine with the AV’s short-term intent to produce an action such as braking, cruising, or changing lanes.Facts include speed, lane, relative distance, predicted trajectories, road structure, and traffic controls.Source: §6 · p. 4

6.1 Driving Rules

• The Driving Rules Catalog groups 35 rules by action, using defaults to suggest actions and exceptions to filter out unsafe or impossible ones.For example, a lane change may be considered for an obstacle or overtaking, but blocked lanes can prevent it.Source: §6.1 · p. 5, §6.3 · p. 6

6.2 Experimentation & Testing

• Testing scenarios cover normal and adverse conditions, including corner cases, with examples such as braking when surrounded and yielding before a merge.The scenarios also include responding to an object entering the lane by changing lanes if possible or braking.Source: System Testing · p. 5, Figure 5 · p. 5
• Half a second on average per frame; AUTO-DISCERN also reached correct decisions in every tested case where an ML-based system failed.The KITTI runtime depends largely on the number of objects and the complexity of the action.Source: Common Driving Scenarios · p. 6
• In KITTI examples, AUTO-DISCERN waited for a clear left lane before merging and delayed a right turn when predicted object trajectories made it unsafe.The scenarios draw on examples from the driving-rules catalog.Source: Common Driving Scenarios · p. 6

6.3 Discussion

• s(CASP) proof trees justify driving decisions through rules; the authors also argue that reasoning over inputs can help infer a consistent view of the world for safer decisions.Source: §6.3 · pp. 6–7
• Commonsense rules can sanity-check a possibly misread speed limit against reasonable speeds for the surroundings and prioritize sensor information when it conflicts with visual information.The paper gives the example of a 35 speed limit being read as 85.Source: §6.3 · p. 7
• Rule-based reasoning can accommodate changed driving conventions across countries, while more nuanced commonsense reasoning and complex temporal scenarios remain areas for further development.The paper notes that differing traffic signs may still require retraining the deep-learning scene-understanding system.Source: §6.3 · p. 7

7 Related Work and Conclusions

• Earlier ASP driving work supported visual sense-making, while AUTO-DISCERN uses commonsense reasoning to understand scenes and compute driving decisions [Suchan et al. 2018; Suchan, Bhatt, and Varadarajan 2019].The cited work takes visual observations from deep learning and provides visuo-spatial semantics at each timestamp.Source: §7 · p. 7
• Unlike ASP work that verifies AV behavior at intersections, AUTO-DISCERN applies commonsense reasoning to actual autonomous driving.Other cited formal-reasoning work addresses unsignaled intersections and handover to a human driver.Source: §7 · p. 7
• AUTO-DISCERN represents driving knowledge as a predicate answer set program executed by s(CASP), using sensor data and scene predicates obtained with machine learning.The authors characterize the system as explainable and provably ethical, and state that commonsense reasoning can compensate for machine-learning scene-processing errors.Source: §7 · p. 7
• The paper demonstrates a complete autonomous-driving decision system built by modeling commonsense reasoning in ASP and s(CASP).The authors identify s(CASP) as a reason an AV system based entirely on commonsense reasoning had not been developed before.Source: §7 · p. 7
• Future work includes refining the infrastructure for CARLA and developing an actual AV deployment with an industrial partner.Source: §7 · p. 7