EbruSubutay

Works

Things I've built

From classical stereo geometry to learned vision-language-action control — plus a few interactive things I make for the web.

Robotics & Perception

My main focus — real-time perception, VLA & control for robots.

Jul 2026

SONIC — Whole-Body Control for the Unitree G1

A whole-body policy that drives the full Unitree G1 humanoid — torso, arms and legs — while staying balanced through dynamic motion.

  • Balanced whole-body motion tracking across the full body
  • Dynamic poses — torso rotation, arm swing, foot lift — without losing balance
  • Live rollout captured across nine sampled poses
unitree-g1 whole-body-control humanoid balance motion-tracking
links coming soon

May 2026 – Present

Humanoid Manipulation with a Learned VLA Policy (Unitree G1)

End-to-end robot-learning pipeline for language-conditioned tabletop manipulation (reach / grasp / pick / place) on a Unitree G1 humanoid in MuJoCo.

  • Scripted scenario & demo generation in MuJoCo
  • ACT (Action-Chunking Transformer) policy training in PyTorch
  • Offline + closed-loop evaluation
unitree-g1 mujoco act pytorch imitation-learning
links coming soon
Franka Emika Panda arm in a MuJoCo simulation performing a vision-language-action manipulation task

Nov 2025 – Dec 2025

Vision-Language-Action Control of a Robot Arm (Franka, MuJoCo)

MuJoCo simulation of the Franka Emika Panda arm integrated with a Transformer-based VLA (OpenVLA, quantized) for closed-loop control from camera images.

  • Full perception → action loop; motion generation (IK + PID)
  • Configurable tasks, local inference, logging/benchmarking (also evaluated SmolVLA)
  • Complete simulation-to-policy-evaluation pipeline for embodied foundation models
openvla smolvla mujoco franka ik/pid quantization
View repo ↗

Feb 2026 – Mar 2026

Industrial Video Understanding for Action Recognition

Industrial video-analysis pipeline for robotic task sequences with structured offline evaluation.

  • Preprocessing, frame selection, temporal redundancy handling
  • Compared baseline modeling choices for action recognition
  • Analyzed prediction stability, confidence trends, failure cases for deployment-oriented CV
video-understanding action-recognition temporal-modeling evaluation
View repo ↗
Animated stereo depth pipeline output showing disparity map and estimated object distance overlays

Dec 2025 – Jan 2026

From 2D Perception to 3D Object Understanding: Bounding Boxes vs. Segmentation

Real-time stereo depth pipeline (SGBM and CreStereo) converting disparity to metric 3D geometry via reprojection and triangulation.

  • Disparity → metric depth → 3D geometry (reprojection / triangulation)
  • Bounding-box vs. segmentation-based lifting of 2D objects to 3D
  • Robustness under viewpoint and occlusion changes
stereo sgbm crestereo 3d-vision depth geometry
Read on Medium ↗ View repo ↗

Interactive & Web Apps

Beyond robotics — interactive tools & data-viz I build for the web.

Physics Atlas web app: an interactive knowledge graph of physics topics on a starfield background, with prerequisite edges connecting concept nodes

Side project · Ongoing

Physics Atlas — Learn Physics by Prerequisites

An interactive prerequisite map of physics — explore how concepts build on each other, from high-school algebra up to quantum gravity, and generate a learning path to any goal topic.

  • Interactive knowledge-graph of ~40 physics topics with prerequisite edges (click a concept to light its full dependency tree)
  • "Learning goal" mode builds a path to any target topic; live search + full-map view
  • Nodes categorized (foundation math · core physics · advanced · goal) over a starfield UI
react vite cytoscape knowledge-graph data-viz education github-pages
Visit live ↗
More on GitHub ↗