Online adaptation of the recovery-switching safety value: far-goal safety improves, everyday safety degrades, and the trade-off reproduces across two independently selected base policies.
Read more
Reproduced and analyzed a legged-robot reinforcement-learning implementation using Isaac Gym and PyTorch. Investigated latent-policy sensitivity and the training-to-deployment velocity-input mismatch, implemented a targeted training ablation, and evaluated its limitations through controlled simulator experiments.
Read moreAn implementation exploring decentralized coordination, spatial organization, and multi-agent consensus algorithms for networked robotic units.
Simultaneous Localization and Mapping on a mobile differential drive platform utilizing pure visual feature odometry, breaking dependency on traditional LiDAR hardware.
A real-time game data parser designed to ingest high-frequency racing telemetry streams and map data straight into physical servo mixes and custom driver modules.