Master thesis - Embodiment-Aware Autonomous Driving Policies with Deep Learning

30 credits – Embodiment-Aware Autonomous Driving Policies with Deep Learning 

Background  

Recent advances in autonomous driving have been largely driven by large-scale imitation learning (IL) from human driving data. By leveraging large datasets of expert demonstrations, learned driving policies have achieved impressive performance across a wide range of driving scenarios. A key challenge for autonomous trucking is that vehicle configurations differ substantially in geometry, articulation, axle layout, payload capacity, braking authority, and steering constraints. These differences influence feasible driving behaviors and motion-planning decisions. While separate policies can be trained for each vehicle configuration, such an approach scales poorly as fleets become increasingly diverse. Conversely, learning a shared policy across multiple vehicle types is challenging because embodiment-specific behaviors must be preserved while still leveraging common driving knowledge. 

Similar challenges have recently been studied in robotics under the umbrella of multi-embodiment learning. Prior work has shown that conditioning policies on robot morphology, kinematics, or learned dynamics representations can improve transfer across heterogeneous platforms1,2. These ideas suggest that embodiment-aware learning may provide a scalable approach for developing autonomous driving policies that generalize across diverse vehicle configurations while adapting to their unique constraints. 

Objective 

Inspired by related works3, this thesis investigates how vehicle-configuration-aware driving policies can be learned from heterogeneous driving data. In particular, the work will explore how policies trained on passenger-vehicle data can be adapted to heavy-duty vehicle configurations, whether embodiment conditioning improves generalization across vehicle platforms, and how training on diverse multi-vehicle datasets influences policy performance and behavior. 

Job description 

The goal of this thesis is to investigate the learning of vehicle-configuration-aware driving policies using imitation learning and embodiment conditioning. The assignment is divided into several subtasks: 

• Develop or extend a motion-planning policy capable of incorporating vehicle-configuration information.  

• Evaluate transfer-learning strategies for adapting policies from passenger vehicles to heavy-duty vehicle configurations.  

• Investigate how driving data from multiple vehicle configurations can be combined within a unified learning framework.  

• Implement, train, and evaluate the proposed solutions in terms of their generalization capabilities, data efficiency, and driving performance. 

Education/program/focus 

Master (civilingenjör) in computer science, robotics, electrical engineering, or applied mathematics, preferably with specialization in artificial intelligence algorithms.  

Number of students: 1-2  
Start date: January 2027    
Estimated time needed: 20 weeks  

Contact persons and supervisors 

Caroline Skoglund, Industrial PhD Student (KTH Robotics, Perception and Learning) caroline.skoglund@scania.com 

Yunus Emre Sahin, Research & Development Engineer in Autonomous Motion yunus-emre.sahin@scania.com