Puru Ojha

I'm a Master's student at the International Institute of Information Technology, Hyderabad. My research interests include Robot learning for long-horizon manipulation tasks, with a focus on zero-shot policy transfer, language-grounded perception, and developing neuro-symbolic systems that combine the generalization of deep learning with the verifiability and safety of formal methods.

Email  /  LinkedIn  /  Github

Research

I'm interested in robot learning for long-horizon manipulation tasks, with a focus on zero-shot policy transfer, language-grounded perception, and developing neuro-symbolic systems that combine the generalization of deep learning with the verifiability and safety of formal methods.

Engineered a framework for procedural generation of complex, structured environments for benchmarking RL agents and studying sim-to-real transfer of navigational policies.

Research Projects

Developed a zero-shot policy transfer framework for cross-embodiment manipulation, enabling a generalist policy (π₀) trained on Franka Panda data to control a morphologically distinct uFactory xArm without any fine-tuning. Implemented a Brownian bridge diffusion model for visual domain adaptation, translating sensory inputs from the target robot (xArm) to the source domain (Panda) for seamless, real-time policy execution. Engineered a robust policy transfer pipeline that leverages real-world data, addressing generalization challenges across different robot hardware and environments.

Designed a graph-based transformer model for predicting object removal sequences in cluttered tabletop scenes, conditioned on natural language goals. Built a large-scale simulation and training pipeline with expert A* supervision and PPO fine-tuning, achieving state-of-the-art performance over GRN and NRP baselines.

Developed a transformer framework for grounded situation recognition to enable robots to follow complex natural language instructions. Improved noun grounding and verb prediction by integrating CLIP and Faster-RCNN features into a CoFormer architecture, achieving state-of-the-art performance on the SWiG dataset.

Engineered a framework for procedural generation of complex, structured environments for benchmarking RL agents and studying sim-to-real transfer of navigational policies. Designed a graph-based algorithm to generate scalable, architecturally consistent layouts with dynamic multi-path structures for realistic simulation.

Miscellanea

Feel free to steal this website's source code. Do not scrape the HTML from this page itself, as it includes analytics tags that you do not want on your own website — use the github code instead. Also, consider using Leonid Keselman's Jekyll fork of this page.