Robot Learning Researcher

[email protected] | LinkedIn | GitHub

Computer Engineering graduate from Cukurova University (2025) specializing in robot learning, reinforcement learning, and humanoid control. Building hierarchical control systems that combine vision-language understanding with low-level motor control for humanoid robots.

RL-to-IL-to-VLA pipeline for the G1 humanoid. Collecting expert demonstrations from trained RL policies, then distilling into end-to-end visuomotor policies.

• Pipeline: RL expert rollout -> LeRobot v2 dataset -> ACT/Diffusion Policy -> GR00T N1.6
• Data: 22-dim joint actions (15 loco + 7 arm), RGB observations from tiled camera
• Goal: End-to-end vision-to-action without hand-crafted skill primitives

End-to-end hierarchical system for Unitree G1 humanoid: VLM task planning + RL locomotion + arm manipulation + drawer interaction.

• High-level: VLM planner (Qwen3-VL) generates skill sequences from natural language tasks
• Low-level: PPO-trained locomotion (17k steps/sec) + arm policy (3cm precision)
• Skills: Walk, reach, grasp, pull drawer, pick-and-place -- 6/6 drawer, 8/8 pick-place success
• Stack: Isaac Lab, RSL-RL, PyTorch, Ollama, CUDA

Multi-stage PPO training pipeline for Unitree G1 whole-body locomotion. Flat walking, velocity tracking, terrain adaptation, torso stabilization, and arm coordination.

• 5-stage curriculum: Flat -> Velocity -> Terrain -> Torso -> Arm control
• Performance: 17k+ steps/sec, 4096 parallel envs, L7 curriculum
• Stack: Isaac Lab, RSL-RL, PyTorch, CUDA

PPO from scratch for quadruped locomotion in NVIDIA Isaac Lab. 17k+ steps/sec on RTX 5070 Ti, 4096 parallel envs, domain randomization for sim-to-real.

From-scratch PPO and SAC for MuJoCo Ant-v5. Pure NumPy/PyTorch, custom reward shaping, 2700+ reward in 8M steps.

Benchmarking ROS navigation planners (TEB, DWA, MPC, Lattice) on BARN dataset. Under publication review.

Interactive Streamlit app for real-time RL training visualization on CartPole.

Robot Learning: PPO, SAC, GAE, Domain Randomization, Curriculum Learning, Flow Matching, Imitation Learning Simulation: NVIDIA Isaac Lab/Sim, MuJoCo, Gazebo, ROS/ROS2 AI/ML: PyTorch, YOLO (v4-8), OpenCV, VLMs (Florence-2, Qwen3-VL) Platforms: Unitree G1, Jetson Nano, Pixhawk Languages: Python, C/C++, CUDA

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

• Led 10+ member team for autonomous UAV systems at TEKNOFEST
• YOLOv7 + Jetson Nano + Pixhawk integration for real-time detection

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