Theses

The two long write-ups behind my degrees, as submitted.

'26

MSc dissertation · University of Edinburgh

Manipulation in the Dark: Blind Sim-to-Real Transfer of a Baoding Policy on the Shadow Hand Lite

Ayush Deshmukh — supervised by Elle Miller and Prof. Sethu Vijayakumar, Statistical Learning and Motor Control Group, School of Informatics, 2026

I took a blind Baoding policy (joint readings and binary touch only) from simulation onto the real Shadow Dexterous Hand Lite with no hardware training data. I split the sim-to-real gap into proprioception, tactile and dynamics parts and closed each one, and the final policy did 55 full 180° rotations in one run without a drop, up to six per 10 seconds. The dissertation also covers the tendon-slack investigation and three attempts at learning an actuator network.

Follow-up paper

'25

B.Tech major project · NIT Karnataka, Surathkal

Reinforcement Learning Powered Generative Models for User-Centric Recommendations

Ayush Deshmukh, Siddharth Kelkar — supervised by Dr. Anand Kumar M, Department of Information Technology, April 2025

Recommending sequences of games with RL: a generator proposes game sequences, a user-preference evaluator scores them, and PPO trains the generator on that feedback. Games are represented with CLIP and BART embeddings of their art, titles and descriptions. Experiments showed that the dual-discriminator setup, full metadata and multimodal content each beat their simpler versions. This became the COMSNETS 2026 paper.

PDF Project page COMSNETS paper