Mihajlo Stevanović
MSc MVA student · ENS Paris-Saclay · Robot learning
Paris region, France
Open to research internships · spring 2027 · get in touch
I'm interested in how robots can learn control from data rather than having it hand-tuned, through imitation and reinforcement learning, and world models that let them plan.
- NowMSc MVA, ENS Paris-Saclay · French Government Scholar (BGF)
- ResearchIL/RL policies for cable insertion in the Intrinsic AI for Industry Challenge, ETF Robotics Lab
- ThesisACT vs. Diffusion Policy on 30k demos: lateral error is the main obstacle; more data cut it 3.5× (41.2 → 11.8 mm)
- World modelsAction-conditioned video diffusion on BAIR robot data: the arm moves in the commanded direction 84.8% of the time
- ControlModel-predictive control for intercepting a thrown ping-pong ball with a Franka Panda (in progress)
- BackgroundBSc Electrical Engineering, University of Belgrade · GPA 9.11/10
- imitation learning
- reinforcement learning
- RL fine-tuning
- world models
- planning
- contact-rich manipulation
- sim-to-real
news
| Sep 28, 2026 | Started classes in the MVA master’s at ENS Paris-Saclay as a French Government (BGF) scholar. |
|---|---|
| Aug 16, 2026 | At PSIML 2026 (Aug 5–16) I built an action-conditioned video world model for robot manipulation. |
| Jul 15, 2026 | Graduated with a BSc in Electrical Engineering from the University of Belgrade (GPA 9.11/10). Thesis: limits of imitation learning for high-precision insertion. |