AI Research Scientist · PhD Candidate

Bojan Derajić

Learning-based safe robot navigation in unknown and dynamic environments

Download Resume
Portrait of Bojan Derajić

About

I am an AI Research Scientist at the AI Lab Berlin of AUMOVIO (formerly Continental Automotive), and a PhD candidate at the Technical University of Berlin as part of the IMRC Lab. My research focuses on safe motion planning and control of mobile robots operating in dynamic and unknown environments — settings where both classical guarantees and data-driven flexibility are needed to make a system trustworthy.

In my approach I bring together control theory and machine learning: Hamilton-Jacobi reachability, model predictive control, control barrier functions, hypernetworks, and neural constraints. Before my PhD, I obtained a MSc degree from the University of Belgrade (thesis on adaptive nonlinear MPC for quadcopters) and a BSc at the University of Banja Luka (thesis on trajectory planning for robot manipulators). Alongside my research, I have worked as a Research and Teaching Assistant and have filed five patent applications, four of them as first inventor.

Selected Publications

Full list on Google Scholar →

  1. VertexCBF: Improving Neural Control Barrier Functions via Vertex-Restricted Control Search

    B. Derajić, S. Bernhard, W. Hönig

    Conference on Robot Learning (CoRL), 2026 · accepted

    arXiv · GitHub

    Figures
    Inverted pendulum trajectories and labels generated by full-control MPPI (left) versus vertex-restricted beam search (right)

  2. CN-CBF: Composite Neural Control Barrier Function for Robot Navigation in Dynamic Environments

    B. Derajić, S. Bernhard, W. Hönig

    IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2026

    arXiv · GitHub

    Figures
    Structure of the CN-CBF: relative-state encoding feeds a neural network whose output is composed with the failure function
    Block diagram of the CN-CBF safety filter with robot dynamics, CN-CBF, backward differentiation and QP solver

  3. ORN-CBF: Learning Observation-conditioned Residual Neural Control Barrier Functions via Hypernetworks

    B. Derajić, S. Bernhard, W. Hönig

    IEEE International Conference on Robotics and Automation (ICRA), 2026

    arXiv · GitHub

    Figures
    Architecture of the ORN-CBF safety filter: observation to SDF, hypernetwork parametrizing the main network, SDF interpolation, and CBF-QP

  4. Residual Neural Terminal Constraint for MPC-based Collision Avoidance in Dynamic Environments

    B. Derajić, M.-K. Bouzidi, S. Bernhard, W. Hönig

    Conference on Robot Learning (CoRL), 2025

    PMLR · GitHub

    Figures
    RNTC-MPC framework: object tracking and motion prediction, SDF sequence computation, hypernetwork, and MPC with residual neural terminal constraint

  5. Learning Maximal Safe Sets Using Hypernetworks for MPC-Based Local Trajectory Planning in Unknown Environments

    B. Derajić, M.-K. Bouzidi, S. Bernhard, W. Hönig

    IEEE Robotics and Automation Letters (RA-L), 2025

    IEEE Xplore · GitHub

    Figures
    Architecture of the NTC-MPC local planner: SDF from sensor data feeds a hypernetwork that parametrizes the terminal-constraint network inside the MPC
    Training diagram: SDF dataset and state grid as inputs, HJ reachability labels, hypernetwork optimized via the loss

Education

Experience

Skills

Programming

  • Python
  • C / C++
  • MATLAB

Tools & Frameworks

  • PyTorch
  • JAX
  • ROS / ROS 2
  • Gazebo
  • CasADi
  • acados
  • MLflow
  • Docker
  • Git
  • Linux

Professional

  • Problem Solving
  • Scientific Writing
  • Time Management
  • Critical Thinking

Awards & Honors