KAIST Mobility Ph.D. Student

Kailang Ma 麻开朗

World Models & Autonomous Driving

I study how intelligent systems can understand dynamic environments, anticipate decision-relevant consequences, and act reliably under uncertainty.

Email Google Scholar GitHub
World Models Autonomous Driving Embodied AI Trustworthy ML
Portrait of Kailang Ma
Updates

News

Selected academic and professional milestones.

Published iLRG at the International Conference on Learning Representations.

Profile

About & Research

Building learning systems that reason about dynamic worlds and remain dependable when conditions change.

I am a Ph.D. student at the KAIST Cho Chun Shik Graduate School of Mobility, where I study world models and autonomous driving. Before joining KAIST, I worked as a Perception Algorithm Engineer on the Scene Team at WeRide, focusing on traffic-light perception and perception-based monitoring of operational design domain (ODD) conditions.

I received an M.S. in Cyberspace Security from the School of Cyber Science and Technology at Beihang University (BUAA) in 2024 and a B.S. in Information Security from the same school in 2021. My current interests span world models, autonomous driving, embodied AI, and trustworthy machine learning.

01

World Models

Predictive representations of complex, evolving environments.

02

Autonomous Driving

Perception and decision-relevant scene understanding.

03

Embodied AI

Agents that perceive, anticipate, and act in the physical world.

04

Trustworthy ML

Uncertainty-aware systems with reliable and recoverable behavior.

Industry

Experience

End-to-end perception development, from data and learning systems to production-facing C++ components.

Perception Algorithm Engineer · Scene TeamWeRide · Perception Team
–
  • Developed traffic-light detection and recognition systems and perception models for monitoring ODD-relevant conditions, including weather, sensor operating status, and road-surface conditions. The resulting signals supported decisions on whether prevailing conditions permitted Robotaxi operation within the defined ODD.
  • Designed and implemented methods for image classification, state classification, 2D object detection, and anomaly detection, contributing across the complete development lifecycle: data management, model training, preprocessing, post-processing, and C++ fallback components.
Perception Algorithm InternWeRide
–
  • Designed a human-in-the-loop framework for semi-automated annotation. At each iteration, a randomly sampled subset of the data pool was manually annotated to estimate the model's precision–recall curve and calibrate confidence thresholds against target precision and recall requirements.
  • Added pseudo-labels that satisfied the calibrated thresholds to the training set, retrained the model, and repeated the cycle. The framework met its design objectives and increased annotation throughput by more than 5×.
Research InternDepartment of Precision Instrument · Tsinghua University
–
  • Contributed to the development and optimization of deep-learning models for wafer defect detection.
Research Output

Selected Publications

Work on safety-critical world models, federated learning privacy, and interpretable deep learning.

Conceptual architecture of the Risk-Informed World Model
Under ReviewarXiv 2026JICV · Q1Perspective
Rethinking World Models for Safety-Critical Embodied SystemsKailang Ma, Heye Huang, Inhi Kim, Kitae JangJournal of Intelligent & Connected Vehicles (JICV, Q1) · Under review

Introduces the Risk-Informed World Model (RIWM), a decision-centric framework that connects decision-relevant representation, counterfactual reasoning, safety-critical episodic memory, and runtime safety assurance.

arXivPDF
Six capabilities for safety-critical embodied world models: generation, modeling, understanding, reasoning, decision, and verification
In PreparationSurvey
World Models for Safety-Critical Embodied Systems: From Generative Simulation to Cognitive Decision-MakingHeye Huang, Kailang Ma, et al.Survey manuscript in preparation · May 2026–Present

Develops a safety-oriented taxonomy spanning representation, prediction, action conditioning, physical grounding, risk awareness, and self-improvement, and reframes evaluation around decision utility, critical-risk recall, counterfactual reasoning, uncertainty calibration, and safe action.

Illustration for iLRG
ICLR 2023First Author
Instance-wise Batch Label Restoration via Gradients in Federated LearningKailang Ma, Yu Sun, Jian Cui, et al.International Conference on Learning Representations

An analytical, instance-wise label restoration method that derives a linear system from gradients and model probabilities for one-step recovery.

PaperCodeSlides
Illustration for Cross-CAM
KSEM 2022Student First Author
Cross-CAM: Focused Visual Explanations for Deep Convolutional Networks via Training-Set TracingYu Sun, Kailang Ma, Xuanxin Liu, et al.International Conference on Knowledge Science, Engineering and Management

A training-set-aware visual explanation method combining gradient and prototype-similarity weights, with Intersection-over-Self (IoS) for measuring target focus.

Paper
Background

Education

Academic training across intelligent systems, security, and machine learning.

Recognition

Competitions & Honors

Selected distinctions in research, autonomous driving, and academic achievement.

2026National Second Prize, 4th NSF OnSite Autonomous Driving Algorithm Challenge — Safety of the Intended Functionality (SOTIF) TrackTeam Leader
2023First Prize, WeRide Excellent Project Award — Traffic-Light Recognition and Prediction
2023Second Prize, ByteDance Security & Risk Control Bootcamp
2021–24First-Class Graduate Academic Scholarship (2022–2024) · First-Class Graduate Entrance Scholarship (2021–2022), Beihang University
2021Second Prize, National Finals, China Collegiate Computer Contest — AI Creativity Challenge (2021 Edition)ProjectCompetition
2017–19National Encouragement Scholarship (2017–2018, 2018–2019) · Second-Class Scholarship for Outstanding Undergraduate Students (2017–2018), Beihang University