Ömer Tarık Banuş

I am a robotics engineer at ASELSAN and hold an MSc in Robotics from Middle East Technical University. I work on how learning and dynamics-based control can work together to help robots move and interact skillfully in the physical world, and on the software that takes these methods from simulation to hardware. My work spans legged locomotion, whole-body control, state estimation, and autonomous drones; I am also interested in dexterous manipulation.

Publications

ICRA 2027 · Under review

Learning Structured Locomotion References for Optimization-Based Whole-Body Control

A reinforcement-learning policy adapts the references of a whole-body impulse controller. In simulation it reached 77.29% mean survival under velocity disturbances, compared with 50.35% for the model-based reference controller.

Thesis architecture diagram: a reinforcement-learning agent sends references to a model-based controller acting on a MuJoCo robot model.

MSc thesis · METU 2026

Learning Hierarchical Task-Space Control Policies for Whole-Body Locomotion

A reinforcement-learning policy decides body motion, stance forces, foothold corrections, and gait timing for a QP-based whole-body controller. Trained in about one hour on 256 parallel environments with a custom C++/Python framework, it recovered from larger horizontal pushes than an MPC+WBC baseline, which remained stronger against vertical ones.

Solo-12 quadruped with the string and pulley loading setup used in the experiments.

IEEE CASE 2024

Contact Consistent Disturbance Estimation for Quadruped Robots

Estimates external body forces and ground reaction forces from proprioception. Evaluated in simulation and on a Solo-12 robot, for which I ported our in-house locomotion controller.

DOI

Projects

ASELSAN · 2022–2025

Blind locomotion on an in-house quadruped

Part of a small team that developed ASELSAN's in-house quadruped from its initial concept, within the ASELSAN–METU legged-robotics collaboration. I contributed to locomotion control and was a key contributor to the C++ software architecture on real-time Linux. My work included ground-plane estimation and slope-aware MPC, stair-climbing swing trajectories, probabilistic contact estimation from joint torques, actuator-aware whole-body impulse control, and the robot's MuJoCo simulation. Without vision, the robot traversed uneven terrain, stairs, and slopes up to 45 degrees.

ASELSAN · 2025–present

Agile autonomous drone

I own the drone's software architecture, developed from project start, and lead GNSS-denied state estimation: a custom error-state Kalman filter and a model-based estimator that accounts for actuator dynamics and aerodynamic effects. I also work on trajectory optimization and Lyapunov-based attitude control, and built the Gazebo and Isaac Sim environments and the software- and hardware-in-the-loop setups used to test the stack before flight.

ASELSAN · System technical manager

Human-following quadruped on Ghost Vision 60

A quadruped that follows a person on foot. It estimates the leader's position from UWB ranging, IMU, and magnetometer measurements with a walking-motion model, and uses LiDAR-based local mapping with Nav2 for routing and obstacle avoidance in ROS 2. I led the system from requirements to field delivery while working directly on tracking and navigation. Our team delivered it in five months and demonstrated it at IDEF 2025.

Bilkent University · 2021–2023

Modelling a collision-resilient micro quadrotor

Flight and landing-compliance dynamics of a foldable, collision-resilient quadrotor. Built its CAD and URDF models, implemented cascaded PID flight control in RotorS, and wrote a custom Gazebo plugin for the landing-gear compliance.

CV

Education

MSc in Robotics, Middle East Technical University (METU), Center for Robotics and AI (ROMER) Sep 2023 – Sep 2026

Thesis: Learning Hierarchical Task-Space Control Policies for Whole-Body Locomotion. Advisor: Assoc. Prof. M. Mert Ankaralı · Co-advisor: Assoc. Prof. Ali Emre Turgut · Collaborator: Prof. Uluç Saranlı.

  • Designed a hierarchical quadruped controller in which a reinforcement-learning policy makes decisions about body motion, stance forces, foothold corrections, and gait timing, while a QP-based whole-body controller computes contact-feasible joint torques.
  • Built a custom training framework with a Python Gym interface for Stable-Baselines3, EnvPool-based parallelization, and a modular C++ control backend embedded in a custom MuJoCo environment; trained the thesis policy in about one hour on 256 parallel environments.
  • Designed the framework to support simulation and hardware deployment, with interchangeable controllers, robot models, and estimators, and a multi-module control architecture with finite-state-machine coordination.
  • Implemented Cheetah Software's model-based locomotion controller and a joint-level reinforcement-learning baseline based on Rudin et al. within the same framework for comparison.
  • Evaluated simulated push recovery against an MPC+WBC baseline, identifying improved horizontal recovery but weaker vertical recovery; investigated individual policy contributions through action-channel ablations in a first-author ICRA 2027 submission.
BSc in Mechanical Engineering, Bilkent University Sep 2017 – Jun 2023
Full scholarship with monthly stipend; honor certificates; four graduate-level electives in control and robotics.

Experience

Specialist Robotic Systems and Control Engineer, ASELSAN Dec 2022 – present

Legged robotic platform Dec 2022 – Oct 2025

  • Worked within the ASELSAN–METU legged-robotics collaboration, through which I conducted my MSc research at METU's Center for Robotics and Artificial Intelligence (ROMER).
  • Developed an in-house quadruped from initial concept as part of a small team; contributed to locomotion control and was a key contributor to the C++ software architecture on real-time Linux.
  • Developed ground-plane estimation, modified convex MPC to account for terrain inclination, and designed foot-swing trajectories for stair climbing; the robot traversed uneven terrain, stairs, and slopes up to 45 degrees without vision on hardware.
  • Used spatial equations of motion to develop probabilistic contact estimation from torque-derived ground-reaction forces; triggered MPC updates when foot-contact states changed.
  • Applied spatial dynamics in whole-body impulse control (WBIC); measured actuator torque limits and incorporated actuator-aware constraints into the QP-based controller to limit knee-torque spikes.
  • Built the robot's MuJoCo simulation and CAD-to-MJCF conversion pipeline.
  • Deployed and tested locomotion controllers on physical quadrupeds; diagnosed failures through logs and hardware experiments and investigated differences between simulation and physical behavior.
  • Helped develop and implement the estimator for our IEEE CASE 2024 paper and deployed the in-house controller on Solo-12 for physical experiments.

Human-following quadruped, Ghost Robotics Vision 60

  • Served as system technical manager from requirements to field delivery; coordinated partner departments and customer discussions while contributing directly to tracking and navigation. Our team delivered the system in five months and demonstrated it at IDEF 2025.
  • Developed leader-position estimation using UWB ranging, IMU, and magnetometer measurements with a walking-motion model; built LiDAR-based local mapping and integrated Nav2 routing and obstacle avoidance in ROS 2.

Agile autonomous drone Oct 2025 – present

  • Own the drone's software architecture, developed from project start; built the ROS 2 architecture and system state machine; coordinate the deployment environment with other teams.
  • Lead GNSS-denied state estimation using a custom error-state Kalman filter; develop a model-based estimator accounting for actuator dynamics and aerodynamic effects, and work on trajectory optimization and Lyapunov-based attitude control.
  • Serve on the product's systems engineering team, coordinating work across teams and making system-level technical decisions; integrate and validate algorithms on prototypes through hardware and flight tests.
  • Developed Gazebo simulations and photorealistic Isaac Sim environments; built simulation-in-the-loop and hardware-in-the-loop test setups, Docker images, and deployment tools so colleagues can run and test the software.
Undergraduate Research Assistant, Bilkent Miniature Robotics Laboratory Oct 2021 – Jun 2023
  • Modeled the flight and landing-compliance dynamics of a foldable, collision-resilient quadrotor.
  • Built CAD and URDF models, implemented cascaded PID flight control in RotorS, and developed a custom Gazebo plugin to simulate landing-gear compliance using C++ and Python.
Flight Controls Engineering Intern, Turkish Aerospace Industries (TAI) Jun – Jul 2022
Performed frequency-sweep system identification for flight-controller analysis.
Robotics Intern, OPLOG – Tarqan Robotics Feb – Jun 2022
Worked on mobile-robot control and state estimation for robotic fulfillment systems.
Teaching Assistant, Bilkent University, Department of Physics Sep 2019 – Jun 2021
Supported laboratory instruction and evaluated students' technical reports and assessments.

Technical skills

Programming
C++, C, Python, MATLAB/Simulink
Robotics software
ROS, ROS 2, Nav2, PX4-Autopilot, MIT Cheetah Software, real-time Linux, Docker, Git, CAN
Dynamics, control, and estimation
Spatial dynamics, model predictive control, QP-based whole-body control, gait and footstep planning, contact estimation, error-state Kalman filtering, sensor fusion
Learning
Reinforcement learning (PPO), PyTorch, Stable-Baselines3, Gym interfaces, EnvPool
Simulation and testing
MuJoCo, Isaac Sim, Gazebo, RotorS, software-in-the-loop and hardware-in-the-loop testing

Honors, training, and activities

  • Team member, ASELSAN–METU/ROMER legged robotics collaboration, which received the 2025 YÖK University–Industry Collaboration Award.
  • Top 0.01% nationally in Türkiye's 2017 university entrance examination (rank 157 of about 2.5 million candidates).
  • Top 0.13% in quantitative and equally weighted scores of ALES, Türkiye's national graduate-admissions examination, 2023.
  • Participant, 10th International Summer School on Screw-Theory-Based Methods in Robotics, 2024: screw-theory-based methods for robot modelling; spatial dynamics for robot dynamics and control.
  • Volunteer building supervisor, Oy ve Ötesi (election-monitoring organization), 2018–2023.

Languages

Turkish (native) · English (TOEFL iBT 110/120, 2026; IELTS Academic 8.0, 2022) · German (elementary)

Contact

[email protected] GitHub LinkedIn