Narendra Aironi

I build World Models for Embodied Intelligence

Hi, I'm Narendra. My passion is building Embodied Intelligence to elevate humanity. My research is currently focused on JEPA based World Models for long-horizon planning. I've also worked on Self-supervised Learning, Deep RL and theoretical Deep Learning. I'm an undergrad @ IIT Bombay.

Current Work

Efficient Planners for World Models

Building a novel MPC based planner for vision-based world models, for efficient exploration of trajectories in latent space for goal-image conditioned tasks. Supervised by Amir Bar, META (FAIR)

Relevant Links

Hierarchical JEPA for long-horizon planning

Following LeCun's position paper to add a Systems-2 JEPA on top of existing WMs, to perform long-horizon tasks. Also design a scalable training method, for H-JEPA. Personal Project

Cross-modal JEPA in Remote Sensing

Cross-modal JEPA-based architecture, that beats SOTA on satellite image retrieval tasks, even with missing input modalities. Paper under review.

Robo Sapiens

Versatile locomotion of a 6DoF biped in challenging terrain using an RL policy trained from scratch. Currently doing sim2real on a custom built biped.

Blender from Scratch

A 3D Modelling and Animation engine like Blender, built from scratch in vanilla C++ and OpenGl. See the short clip created using my engine.

Rizzume

Trained on 500+ handpicked resumes, to suggest job profiles, future projects and optimize for ATS. 1000+ resumes analyzed in 1st week of going live.

Work Experience

Blog