01 — CAPTURE
Learn from demonstrations.
Use human examples to define a skill. Our initial focus is manipulation: turning demonstrations into policies that can reproduce useful behavior.
ROBOT LEARNING. REAL-WORLD AMBITION.
Learning systems
for the physical
world.
Robots should learn useful skills from experience, and keep building on what they know.
We’re building toward reusable robot skills — learned from demonstrations, adapted through experience, and evaluated with safety in mind.
Founder-led. Research-grounded. Early stage.
SKILL WORKSPACECONCEPT / 001
MANIPULATIONPick & place
HUMAN DEMONSTRATION → ROBOT SKILLILLUSTRATIVE
Start with human demonstrations. Capture the task, the motion, and what success looks like.
OUR TECHNICAL FOUNDATIONImitation learningContinual learningModel-based controlPhysical AI
01 / THE APPROACH
A new task shouldn’t
mean starting over.
Robot behavior is still too often engineered one task at a time. Our focus is a learning workflow that makes demonstrations useful, skills reusable, and evaluation repeatable.
02 — ADAPT
Build on what’s learned.
Explore how robots can acquire new tasks while retaining earlier skills. Continual learning is central to our long-term direction.
03 — VALIDATE
Keep the real world in view.
Evaluate learned behavior against task objectives, physical constraints, and failure cases before considering hardware deployment.
PRODUCT DIRECTION
A robot-learning workspace for robotics teams and research labs: organize demonstrations, develop manipulation skills, and compare policy behavior across tasks.
02 / THE INTELLIGENCE LAYER
Reason about the task.
Learn the motion.
Language models can help make a robotics workflow more useful, from describing a task to understanding why a trial failed.
Planned integration with the Claude API
Task specification
Use Claude to help translate natural-language instructions into structured task descriptions and evaluation criteria.
Experiment analysis
Explore Claude-assisted interpretation of rollout logs, failure cases, and policy comparisons to guide the next experiment.
A researcher’s interface
Build a conversational layer for navigating experiments and documenting results, alongside the underlying learning tools.
The planned architecture keeps language-model reasoning at the task and analysis level. Learned policies and dedicated controllers handle robot motion.
03 / RESEARCH FOUNDATIONS
Grounded in work
on real robots.
Selected projects from the founder’s academic research at TUM. These experiments inform our approach to learning, high-level reasoning, and real-hardware integration.
REAL-HARDWARE RESEARCHREINFORCEMENT LEARNING / CONTROL
Learning to recover.
With safety constraints.
A learned quadrotor recovery policy with a Control Barrier Function filter, deployed on an Agilicious drone.
Explore the researchFOUNDATION MODELS / ROBOTICS
High-level reasoning.
Hardware-level control.
A ROS 2 driving stack combining language and vision models with classical control and a safety state machine.
Explore the research
OGUZHAN ESENFounder
04 / THE FOUNDER
Research curiosity.
An engineer’s mindset.
I’m Oguzhan Esen, a robot learning and control engineer based in Munich. Esen Robotics is the venture I’m building around a question: how can robots keep learning useful skills?
My background combines robot-learning research at the Technical University of Munich with more than four years of industrial control engineering at AVL. My work spans safe reinforcement learning, generative policies, continual learning, and real-hardware integration.
Esen Robotics is a founder-led, early-stage venture hosted at oguzhanesen.com. This domain is our home on the web and the domain for our business email.
LET’S BUILD THE NEXT CAPABILITY.
What should your
robot learn next?
We’d like to hear from robotics teams, research labs, and potential collaborators. Tell us about your task, your platform, and where you’re getting stuck.