Robot Learning Researcher | Hierarchical VLM-RL for Humanoid Control

mehmetturanyardimci@hotmail.com | LinkedIn | GitHub

I'm a Computer Engineering graduate from Çukurova University (2025) specializing in robot learning, reinforcement learning, and humanoid control. My research focuses on hierarchical VLM-RL systems that combine vision-language understanding with low-level motor control for humanoid robots.

Current Research: Developing a hierarchical control system for the Unitree G1 humanoid robot in NVIDIA Isaac Lab, combining PPO-based locomotion with Flow Matching manipulation policies.

Hierarchical VLM-RL for Humanoid Manipulation

• High-level: Vision-Language Models (Florence-2/Molmo2) for semantic scene understanding
• Mid-level: Flow Matching for bimanual manipulation planning
• Low-level: PPO policies for whole-body locomotion control
• Achieved 17,000+ steps/second training speed with 4096 parallel environments

Robot Learning: PPO, SAC, Actor-Critic, GAE, Domain Randomization, Curriculum Learning
Simulation: NVIDIA Isaac Lab, Isaac Sim, MuJoCo, Gazebo, ROS/ROS2
AI/ML: PyTorch, TensorBoard, YOLO (v4-8), OpenCV, Flow Matching
Platforms: Unitree G1, Jetson Nano, Pixhawk
Languages: Python, C/C++, CUDA

Hierarchical control system for Unitree G1 humanoid robot combining VLM planning with RL execution.

Architecture:

• Stage 1-2: Locomotion policy (velocity tracking, terrain adaptation)
• Stage 3: Torso control (pitch/roll/yaw tracking while walking) — 24.69 reward
• Stage 4: Arm control with residual actions (77 obs dim, 22 actions) — 28.81 reward
• Stage 5+: VLM integration for language-conditioned manipulation

Technologies: Isaac Lab 2.3.1, RSL-RL, PyTorch, CUDA, TensorBoard

PPO implementation from scratch for quadruped locomotion in NVIDIA Isaac Lab.

Highlights:

• 17,000+ steps/second on single RTX 5070 Ti GPU
• Domain randomization for sim-to-real transfer
• Vectorized environments with 4096 parallel instances

From-scratch PPO & SAC implementation for MuJoCo Ant-v5 environment.

Key Features:

• Pure NumPy & PyTorch
• Custom reward shaping to eliminate hopping behavior
• 2700+ reward achieved in 8M steps
• GAE (λ=0.95), learning rate annealing, exploration decay

Benchmarking framework for ROS navigation planners using BARN dataset.

Planners Tested: TEB, DWA, MPC, Lattice
Status: Under publication review at Çukurova University Engineering Journal

Interactive Streamlit app demonstrating real-time RL training on CartPole.

• Watch the agent learn live in browser
• Adjustable hyperparameters for educational purposes

UAV Team Captain | 1.5 Adana AGM ALKAR (3 years)

• Led 10+ member team designing autonomous UAV systems
• Integrated YOLOv7 + Jetson Nano + Pixhawk for TEKNOFEST competitions
• Developed real-time object detection for fixed-wing UAVs

• Hierarchical Control: Whole-body humanoid manipulation with PPO + Flow Matching
• VLM Integration: Florence-2/Molmo2 for semantic scene understanding and task planning
• Sim-to-Real: Domain randomization and teacher-student distillation for G1 deployment

Open to research collaborations and R&D opportunities in humanoid robotics and robot learning.

📧 mehmetturanyardimci@hotmail.com