Mehmet Turan Yardimci

Mehmet Turan Yardimci

Computer Engineering Graduate · Robot Learning Researcher

Building reinforcement learning systems for humanoid robot whole-body control. Currently developing a hierarchical VLM-RL loco-manipulation pipeline for the Unitree G1 humanoid in NVIDIA Isaac Lab.

mehmetturanyardimci@hotmail.com GitHub LinkedIn
Download CV
Latest Paper

Critic Architecture Matters: Dual vs. Unified Critics for Humanoid Loco-Manipulation

Mehmet Turan Yardımcı
ICRA 2026 · Workshop on Reinforcement Learning in the Era of Imitation Learning (RL4IL) · Vienna, Austria
📄 Read PDF 🔗 OpenReview

Research Interests

Robot Learning Reinforcement learning for locomotion and whole-body manipulation
Vision-Language Models VLM-guided task planning for robotic manipulation
Sim-to-Real Transfer Domain randomization and curriculum learning for real-world deployment
Autonomous Systems UAV navigation, path planning, and multi-agent coordination

Isaac G1 Humanoid VLM-RL Loco-Manipulation

Hierarchical reinforcement learning system for Unitree G1 humanoid robot — whole-body control with 43 degrees of freedom and Vision-Language Model integration.

G1 Humanoid Arm Reaching Demo
43 Degrees of Freedom
4,096 Parallel Environments
17K+ Steps / Second
Multi-Stage Curriculum Training Pipeline:
Standing→ Locomotion→ Torso Control→ Dual Arm Reaching→ Grasping→ Loco-Manipulation
  • Triple Actor-Critic architecture: Separate policies for locomotion (legs + waist), arm reaching (shoulders + elbows + wrists), and hand control (DEX3 3-finger dexterous hands)
  • Dual Actor-Critic loco-manipulation: Decoupled locomotion and arm policies trained sequentially with curriculum learning, achieving coordinated walking and reaching
  • Anti-gaming mechanisms: Absolute target sampling, 3-condition reach validation, movement-centric rewards to prevent curriculum exploitation
  • 29-DoF whole-body model: 12 leg + 3 waist + 8 arm + 6 wrist joints, with 14 DEX3 finger joints for grasping
  • Goal task: Squat, pick object from ground, stand up, walk to table, place object in box
NVIDIA Isaac Lab RSL-RL PyTorch PPO Florence-2 CUDA Curriculum RL Domain Randomization
GitHub Repository →

Projects

Anymal-C Quadruped

Isaac Lab Anymal-C Quadruped Locomotion

PPO algorithm implemented from scratch for ANYmal-C quadruped robot. Achieved 17,000+ steps/sec training with domain randomization and reward shaping on RTX 5070 Ti.

Isaac Lab PPO PyTorch CUDA

View on GitHub →
MuJoCo Ant-v5

MuJoCo Ant-v5 PPO from Scratch

PPO and SAC implemented with pure NumPy and PyTorch for MuJoCo Ant-v5. Achieved 2700+ reward by eliminating hopping through custom reward shaping. 16 parallel environments with GAE.

MuJoCo PPO SAC NumPy

View on GitHub →
Actor-Critic CartPole

Live Actor-Critic Training (CartPole)

Interactive Streamlit web app demonstrating real-time RL training. Actor-Critic algorithm with adjustable hyperparameters for educational visualization.

Streamlit Actor-Critic Gymnasium

View on GitHub →
BARN Benchmark

BARN Benchmark: Local Path Planners

Benchmarking framework using ROS and the BARN dataset to evaluate local path planners (DWA, TEB) in cluttered environments. Academic paper pending publication.

ROS Gazebo Python Navigation

View on GitHub →
YOLOv7 UAV Detection

YOLOv7 Fixed-Wing UAV Detection

Real-time object detection system for autonomous UAV operations using YOLOv7, integrated with Jetson Nano and Pixhawk flight controllers for TEKNOFEST competitions.

YOLO OpenCV Jetson Nano ROS

View on GitHub →
NXT Robot PID Control

PID Control with NXT Robot

PID controller implemented on LEGO Mindstorms NXT robots for precise motor control with real-time sensor feedback, using NXC language.

PID NXC Control Systems

View on GitHub →

Education

Oct 2021 — Oct 2025

Bachelor's Degree in Computer Engineering (English)

Cukurova University, Adana, Turkey
  • GPA: 2.85 / 4.00
  • Thesis: Benchmarking Local Path Planners in ROS using the BARN Dataset
  • Focus: AI, robotics, pattern recognition, optimal control
Sep 2024 — Feb 2025

Erasmus+ Exchange — Computer Science (English)

Bialystok University of Technology, Poland
  • GPA: 4.9 / 5.0
  • Focus: Robotics, PID control, HCI, computer graphics

Experience

Aug — Sep 2025

Computer Engineering Intern

Kivanc Tekstil, Adana
  • In-house software development with .NET and C#
Jul — Aug 2025

Computer Engineering Intern

Medcem, Adana
  • Algorithmic solutions for cement production automation
  • Software architecture, debugging, and UI design with .NET
Feb 2022 — Oct 2024

Team Leader

1.5 Adana AGM Alkar UAV Team
  • Led 10+ members designing autonomous UAV systems
  • TUBITAK and TEKNOFEST competition projects
  • AI-based image processing for autonomous navigation (YOLOv5, Jetson Nano, Pixhawk)

Publications

Critic Architecture Matters: Dual vs. Unified Critics for Humanoid Loco-Manipulation
Yardimci, M.T.
ICRA Workshop on Reinforcement Learning in the Era of Imitation Learning (2026)
Benchmarking Local Path Planners in ROS using the BARN Dataset
Yardimci, M.T., Cogurcu, Y.E.
Cukurova University Journal of the Faculty of Engineering (Pending Publication, 2026)

Skills

Robot Learning
PPO Actor-Critic GAE SAC Reward Shaping Domain Randomization Curriculum Learning
Simulation
NVIDIA Isaac Lab MuJoCo Gazebo RViz
AI / ML
PyTorch TensorBoard YOLO OpenCV CNNs Florence-2
Programming
Python C / C++ CUDA Linux Git NumPy ROS / ROS2
Hardware
NVIDIA Jetson Pixhawk RTX 5070 Ti Unitree G1
Certifications
Supervised ML: Regression & Classification (Coursera)