Thirulok Sundar Mohan Rasu

Enhanced the full software stack for a visual-servoing-based auto-fastening system that derives target poses from a single demonstration, enabling CAD-free robotic bolt fastening on parts moving on a conveyor; demonstrated the system to investors, customers, and integrators. Added a detector-free semi-dense matching method and a real-time tracker so pose estimation stays robust through occlusion and low-texture parts, raising vision-stack throughput from 14.5 Hz to 21 Hz via compressed image transport. Built an Isaac Sim camera-mount pose optimization pipeline that renders and scores candidate mounts across CAD parts (90–565 mm) with multi-plane bolts via staged search over 1,000 candidates. Tuned the MPC controller with automatic weight tuning via preferential Bayesian optimization across 30 bolt positions, cutting mean settle time from 4.17 s to 2.67 s (31%) on a FANUC arm. Extended working conveyor speed from 60 to 100 mm/s on a UR10e, cutting cycle time up to 5.9× (21.3 s to 3.6 s) while holding 98% success across 150 runs. Ported the stack to a FANUC CRX10iA-L, resolving driver-level jerk instability to sustain a continuous 159-cycle run, and authored a state machine for multi-plane fastening reaching 95% success across 200 runs.

Evaluating LeWM, a JEPA-based world model, with free-energy and coding-rate regularizers as replacements for the SIGReg anti-collapse regularizer — improving hard-protocol (100-step planning) accuracy on a two-room navigation dataset from 16% to 75% with the coding-rate variant. Showed that a free-energy (Gaussianizing) loss improves reconstruction and classification across CNNs, MLPs, and autoencoders, built a joint Gaussianizing classifier, and published freegaussianizer as an open-source package. Separately, translating the CyScat electromagnetic-scattering simulation codebase from MATLAB to Julia and Python and implementing differentiable simulation via forward-mode automatic differentiation through the T-matrix solver, enabling gradient-based optimization of refractive index and wavelength for maximum wave transmission through photonic structures; validated to machine precision against finite-difference checks.

Engineered an end-to-end data generation and labeling pipeline to curate a 1M+ sample dataset of real/synthetic audio pairs in structured JSON. Fine-tuned the CLAP audio-language model achieving 97% accuracy on deepfake speech detection, outperforming human listeners. Paper submitted to Interspeech 2026 (under review).

Developed Multi-head Mask R-CNN models for instance segmentation and object detection in warehouse environments and integrated them into production robot perception pipelines. Implemented perceptual hashing for large-scale dataset deduplication (30% size reduction) and built order-fulfillment optimization using Leiden clustering.