Rushav Dash

Capstone project: a bio-inspired quadcopter that lands and perches on its own. Instead of relying on an off-board system or a pre-surveyed target, the drone finds and evaluates a landing site with machine vision running on its own compute, then executes the approach and landing autonomously. [TODO: confirm scope — target type, platform, and what "perching" means mechanically.]

• Pitched the project and got it approved as a student-led capstone.
• Assembled the team.
• Began initial design and research with the team.

The qualification task puts a UR5e in Gazebo in front of a task board whose position, orientation, and port layout are randomized every run, and asks it to insert SFP and SC fiber-optic connectors using only three wrist cameras, a force/torque sensor, and joint states. Over ten weeks I built the whole stack from zero: per-port ResNet-18 regressors that locate ports in camera pixels, back-projection through the camera intrinsics to a 3D target, and a Cartesian impedance controller that approaches, aligns, and descends with force-modulated step sizes. A parallel imitation-learning track collected 530 demonstrations (169 GB) and trained an ACT policy with LeRobot on an RTX 5090. I did not achieve a full insertion; the gap between "close" and "inserted" on a 14 mm port from pixels alone is the honest headline.

• Trained eight generations of per-port ResNet-18 detectors (four models, one per port) and fixed the biggest score jump of the project by retraining on 360° yaw data, which added roughly 60 points on hard board configurations.
• Wrote 20+ versions of the classical policy on top of the Cartesian impedance controller, including the force-modulated descent (full step below 5 N, half step 5–10 N, halt above 12 N) and the SFP descent-depth fix that removed a 24-point contact penalty per trial.
• Built the data-collection recorder, the LeRobot/ACT training track, and a multi-config docker-compose benchmark that exposed detector overfitting; shipped 46+ Docker images to the competition portal under the sensor whitelist.

Large companies keep deep expertise in the heads of senior staff, and documenting it is slow and rarely searchable in context. Thoth turns capture into a guided AI interview, storage into a per-subject ChromaDB collection, and retrieval into scoped, grounded answers with mandatory citations. It is an orchestrator, not an answerer: a Claude Haiku 4.5 classifier routes each question by confidence to a subject-scoped Sonnet 4 agent, a clarifying question, or an admin escalation. Knowledge passes a two-stage SME-then-admin approval before it is indexed, and with zero approved entries the query path refuses to call the model at all. Stack: FastAPI, SQLAlchemy + SQLite, ChromaDB, React + Vite + Tailwind.

• Scaffolded the backend (agents, routes, services, models, seed data) and the React frontend, then built the four-role single-window UI and the one-command launcher.
• Built the /api/v1 benchmark surface for the evaluation harness (query, knowledge, health, system endpoints), per-call token telemetry, session handling, and the benchmark test script.
• Owned the subject classifier and confidence routing, tightened the SME answering prompts, and wrote the README, architecture, demo, and production-recommendations docs.

Pick a drink from the menu and two SO-101 arms make it: the right arm handles alcohols, the left handles mixers, and each runs six phases (find, go, align, grab, pour, toss) in sequence so they never share the workspace. An overhead RealSense D435 plus an eight-bottle YOLO model gives a coarse 3D bottle pose from the nearest 10% of depth pixels; at standoff a wrist-camera YOLO drives a proportional visual-servo loop to center the bottle before a TF-computed grasp, and MoveIt 2 plans the collision-aware motion. The team first trained an end-to-end ACT/LeRobot policy, which learned to grab and pour but failed unpredictably with no way to inspect a bad rollout, and pivoted to this explicit per-stage pipeline.

• System orchestration: sequenced the six phases (find, go, align, grab, pour, toss) per arm and the right-then-left hand-off between arms, with a stop-after-stage switch for step-by-step bring-up.
• Wrist-camera detection and visual-servo alignment: a stock COCO YOLO on each wrist camera that accepts bottle or vase and locks onto the largest box, driving a proportional loop on shoulder pan until the bottle is centered before the grasp.
• End-to-end tuning of the full pipeline through the 480-trial evaluation: standoff distance, per-shelf grasp offsets, gripper-axis approach, gravity-sag compensation, and retry-then-continue on the pour and toss.

The TECHIN 516 brief asked for a robot that navigates dynamic indoor spaces and interacts with people; we wanted ours to have personality. ShoeHugger is a modified TurtleBot3 Waffle that detects a person with YOLOv8n-pose, tracks their ankle keypoints through an EMA-smoothed lock, and follows at a set distance while a three-zone LiDAR guard (stop, slow, nominal) keeps it safe. SLAM localization and Nav2 planning run underneath, a custom TF broadcast exposes the person frame, and a third Dynamixel drives a four-bar wing linkage for the Attack and Stalk behaviors. Everything, including the Tkinter control UI, comes up from one launch script.

• Built the person-following perception: YOLOv8n-pose ankle-keypoint tracking at imgsz 256 with an EMA-tracked lock so the foot target stays stable through occlusion.
• Integrated SLAM, Nav2, and a custom TF tree, and wrote the three-zone LiDAR obstacle logic that governs stop, slow-down, and nominal following.
• Designed the Attack/Stalk state machine, the Tkinter control UI over ROS topics, and the single-command launch orchestration for the whole stack.

Eye floaters affect millions of people, and the only current treatment, vitrectomy, is invasive surgery that risks retinal detachment and cataract progression. VitreaClear applies focused ultrasound from outside the eye so that acoustic radiation force displaces floaters out of the visual axis. We characterized transducer beam profiles to set a focal geometry that targets the vitreous chamber without crossing the retinal plane, validated therapeutic parameters (geometry, drive amplitude, pulse duration) over dozens of water-bath runs, and built a real-time closed-loop power controller to hold acoustic output steady across trials.

• Led transducer system design: beam-profile characterization and focal-geometry optimization for the vitreous chamber.
• Developed the real-time closed-loop power controller and the Python/MATLAB signal-processing layer that closes the loop.
• Ran the iterative water-bath validation campaign and assembled the bench prototype.