ROBOTICS AND AI ENGINEER
I’m Enis. I work on humanoid reinforcement learning and legged robot control, and I’m building my own quadruped platform.
UCL Robotics & AI · MSc with Distinction
Open to London and international opportunities · Open to relocation
Figure 1, Yalcin et al., E-SDS (2025).
SELECTED WORK
01 / UCL MSC RESEARCH / 2025
Using video interpretation and terrain sensing to guide automated reward generation for humanoid locomotion.
Reported simulation result; performance varies by terrain and metric.
Explore the research ↓02 / INDEPENDENT / ONGOING
Mechanical design, embedded electronics and model-based control for an independently developed legged-robot platform.
Tested in simulation with sensor noise; physical testing is next.
Explore the platform ↓KING’S COLLEGE LONDON / MSC DISSERTATION
I combined spatiotemporal demand forecasting with electric-vehicle routing for shared-mobility stations across London. The study compared demand coverage and routing outcomes across different station sets.
01 / HUMANOID REINFORCEMENT LEARNING
UCL Robot Perception and Learning Lab
Supervisor: Prof Dimitrios Kanoulas
E-SDS: Environment-aware See it, Do it, Sorted — Automated Environment-Aware Reinforcement Learning for Humanoid Locomotion.
Enis Yalcin, Joshua O’Hara, Maria Stamatopoulou, Chengxu Zhou and Dimitrios Kanoulas.
Accepted at RiTA 2025 · Springer LNNS, in press
UCL publication record ↗Can video-based task information and terrain measurements make automated reward generation more responsive to the environment a humanoid needs to cross?
I developed E-SDS: a vision-language model interprets video and terrain information to generate executable reward functions. PPO training and execution feedback then inform reward refinement. I trained Unitree G1 policies in Isaac Lab and evaluated them on bumps, gaps, obstacles and stairs.
On simple terrain, the reported velocity tracking error fell from 2.225 to 0.387 compared with the manual-reward baseline. E-SDS also learned stair descent in simulation. These are simulation results; transfer to a physical humanoid remains a further validation step.
03 / ROBOT LEARNING
I transferred paths from phone videos into Panda simulation demonstrations and post-trained SmolVLA using mixed data and replay. I tested ordinary pushes, passages and blocking obstacles, alongside a separate world-model action-ranking approach.
79.9% lower target-position error234.6 → 47.3 mm in the paired passage test.
Shown: one paired VLA-only test; world-model ranking was evaluated separately.
Code and full results ↗02 / LEGGED ROBOT DEVELOPMENT
Independent development
Mechanical design · Electronics · Control
I am developing the platform across mechanical design, embedded electronics, communication and control software. The current work focuses on simulation validation and hardware integration.
CONTROL AND ESTIMATION STACK
I have tested the simulated system with sensor noise. In earlier quadruped coursework, I also estimated foot-contact states from proprioceptive data and tested cosine, figure-eight and elliptical path tracking. These experiments made contact uncertainty and controller behaviour concrete before I began the independent platform.
The next step is to integrate the hardware, test the controller on the physical robot and then study how learned policies can improve locomotion without losing the structure of the model-based controller.
SELECTED ACADEMIC PROJECTS
Sensing and manipulation in a Gazebo tabletop environment with a robot arm, parallel gripper and objects of different shapes.
Compared leader–follower, multi-agent path finding and potential-field methods for a drone swarm. The simulation tasks included formation changes, narrow-window traversal, forest-like obstacles and moving obstacles.
Formation split during moving-obstacle navigation
TSP waypoint sequencing, A* planning and ArUco-based arrival verification in ROS2. Evaluated inspection routes across four scenarios.
ArUco target verification from the drone camera
LiDAR-based obstacle detection and ROS2 Bug0/Bug1 local planning, with boundary following, Chaikin smoothing and stuck recovery. Tested in Gazebo and RViz using point-cloud-derived obstacle boundaries.
Estimated four-foot contact states from proprioceptive signals, comparing a learned classifier with hysteresis on held-out quadruped data. A second controller tracked cosine, figure-eight and elliptical reference paths.
Compared COLMAP reconstruction and ORB-SLAM2 on indoor and outdoor image sequences. Examined camera tracking, reconstructed maps and the effects of the scene on localisation.
Implemented 2D state-estimation models and compared EKF-SLAM with G2O factor-graph optimisation under noise, changing landmark range and loop closure. Examined drift, RMSE, covariance bounds and graph-pruning trade-offs.
Built an OpenCV perception pipeline for a rotating target: segmentation and geometric estimation from camera data, then rotation-cycle estimation from video using SIFT feature matching and RANSAC to reject mismatched points.
Applied neural motor-model correction and compared MLP and random-forest trajectory prediction for a simulated 7-DoF Panda arm, including evaluation on unseen targets.
Combined MPC trajectory tracking with EKF localisation for a differential-drive robot. Evaluated landmark layouts, sensor noise, prediction horizons and controller weights.
ABOUT
After studying Industrial Engineering at Sabancı and Engineering at King’s College London, I found my direction in robotics at UCL. My work now connects learned behaviour, model-based control and the practical work of building a robot.
I want to continue exploring robot learning across perception, manipulation and locomotion, including world-action models and how learned behaviour can be tested on physical robots.
MSc Robotics and Artificial Intelligence
Distinction
MSc Engineering with Management
Distinction
BSc Industrial Engineering
Finance minor
RESEARCH AND ENGINEERING OPPORTUNITIES
Interested in research engineer roles and PhD opportunities in robot learning, locomotion and control.
[email protected] ↗