M.Tech student at IIT Madras working across model-based control, machine learning, and computational design. My research spans origami robots, muscle dynamics, policy distillation, and the evaluation of learned systems. Below is a selection of recent and ongoing work.
A coupled dynamic model for a robotic arm that combines motorized joints with pneumatic origami segments. Model-based control tracks 2.5× tighter than a decoupled PID baseline in simulation; distilling the MPC into a neural policy retains 93–94% settled success on the tested goals at much lower decision cost.
Can interpretability improve the distillation of neural control policies into decision trees? Across 27 teachers on four tasks, matched-state augmentation improves response agreement, but ordinary sampling controls recover most of the gain. The study tests what the metric measures and when feature selection removes information needed for control.
Hill-type muscle models with switched activation dynamics and a force–velocity kink do not have one local linearization at equilibrium. They have four directional branches. This study derives the switches and measures how choosing the wrong branch affects prediction and closed-loop control.
Transferring a reversed-comparison test to hard mathematics initially looks like evidence of unfaithful reasoning. Whitespace and rewording controls reveal a different explanation: the model’s stated intermediate answers are sensitive to small prompt changes, without clear evidence of direction-driven rationalization.
Can patterned light act as a synthetic motor pool for engineered muscle? A calibrated simulation finds a wider range of smooth, low force under spatial stimulation, while refuting the original endurance claim. Optical transport bounds how many tissue sectors can be addressed independently.
A re-implementation of Ape-X (Horgan et al., 2018) that replaces the reference
ZeroMQ/asyncio/multiprocessing transport, proxies, and queues with a single
MPI_COMM_WORLD where every logical role is a fixed rank, launched
with one mpirun command. The replay buffer's segment tree is ported
to a C library with optional OpenACC GPU offload, giving roughly two orders of
magnitude speedup, and the MPI port beats the ZMQ baseline on actor intake
throughput and reward growth at every actor count tested.
FDM print quality improves when a layer is deposited as one continuous, crossing-free stroke. Flat-foldable origami tessellations already have the structure this requires: by Maekawa's theorem every interior crease vertex has even degree, so the crease graph is natively Eulerian, and a printable toolpath falls out for free, with no mesh-refinement overhead. Includes a browser-based slicer, a FOLD-format extension for storing toolpaths, and a from-scratch GPU origami simulator.
A from-scratch reimplementation of Ghassaei, Demaine & Gershenfeld's GPU origami simulation method (the engine behind Origami Simulator): a crease pattern becomes a mass-spring mesh driven by four validated force types and an explicit physics loop, with a NumPy reference cross-checked to round-off against a Taichi GPU port that reproduces the paper's ~78× speedup at scale.