i teach machines to see, mostly so they don't run into things.
currently a robotics masters student at WPI, working as a research assistant at the RoboCare Lab on human daily activity learning and recognition basically teaching a robot to notice when you're brushing your teeth vs. brushing the cat. before this i was in india putting flight controllers on drones and a satellite into the sky (it came back down, eventually, with a 2nd place medal attached). on the side i'm building a perception stack i call schrödinger's vision a name i picked because in my experience the lane line is genuinely both there and not there until something's brave enough to look.
location : San Jose, CA / Worcester, ma
status : sleep-deprived but enthusiastic
working on : teaching a humanoid robot to recognize what humans are up to
side quest : multi-camera autonomous driving perception
aspiring to : physical AI, VLAs, and other acronyms i'll explain later
🛣️ MultiCam Autonomous Driving Perception — 4 cameras, 13 driving sequences, 15+ models stitched together, one tesla FSD-shaped dream. detects lanes, vehicles, pedestrians, traffic lights, optical flow, depth, the works. survives night, tunnels, sunset glare, and dense intersections. fails gracefully when it doesn't.
🛰️ CanSat — i co-led a 8-person team that built a satellite the size of a soda can. it had reaction wheels, RF telemetry, sensor fusion, and a healthy fear of gravity. 2nd place nationally at SMOPS 2023 (india's first InSpace CanSat comp). a servo died mid-flight. we adapted. we did not die.
🚁 Crazyflie quadrotor controls — cascaded PD with SO(3) attitude control, polynomial trajectories smooth enough that the math is doing more work than the props. 18+ gain configs swept, RL-tuned, well-damped in 3-4.5s. yes i'm proud of it.
👁️ Visual-Inertial Odometry (deep + classical) — built a CNN+LSTM VIO from scratch and an S-MSCKF for comparison. trained on synthetic blender data, evaluated on EuRoC against vicon. classical still wins on accuracy. deep wins on style points.
🏗️ NeRF & SfM from scratch — positional encoding, hierarchical sampling, volumetric rendering, the whole 3D-is-just-a-function-call deal. plus a full structure-from-motion pipeline because i wanted to know what bundle adjustment actually does.
🔧 OpenManipulator-X — wrote FK/IK by hand, exposed it as a ROS2 service, ran a 7-pose pick-and-place. the robot picked. the robot placed. occasionally in the right order.
📡 GATI — a LoRa-controlled payload deployment system for UAVs. dual servos. microsecond timing. fail-safes. used in the field. did not fall out of the sky.
🔬 [IEEE VLSID 2025] FPGA target ID — fine-tuned YOLOv2-Tiny onto a PYNQ-Z2 board because GPUs are too easy. published the paper. the FPGA still runs.
daily drivers — C++ Python ROS2 PyTorch OpenCV Linux
when the project demands it — C Bash STM32 Raspberry Pi Pixhawk MAVLink Blender PyBullet LaTeX
aspirationally — anything with "physical AI" or "VLA" in the name, ask me again in 6 months
🌐 aryanshah.work · 💼 linkedin · 📧 email · 📸 instagram
if you're a recruiter: hi 👋 i'm looking for summer/fall 2026 roles in perception, autonomy, or embedded robotics. the email above works. the linkedin is faster.
if you're a robot reading this: don't crash.