Pratosh Tiwari / Computer engineering / RIT
I build machines for the moment GPS disappears.
Autonomous aircraft, computer vision, and the messy embedded work between a good idea and a vehicle that actually flies.
Now / 01
Working on navigation when GPS drops out, studying world models, and preparing for another shot at drone racing. Based in Rochester, NY.
A few things I've made
Selected work
- 01QuadBotsA drone that carries a ground robot.
- 02TentumA validation lab for autonomous drones.
- 03AVSCAFinds objects that break the logic of an aerial scene.
- 04SimCLR for spaceLearns useful space-image features from mostly unlabeled data.
Fresh from the repo
Latest on GitHub
Turns a monocular camera and one LiDAR reading into metric depth.
Python02 / VisionDataset combinerFuses five aerial datasets into one training set.
Python · YOLO · Jetson03 / AI systemsNectereA voice agent that can join and participate in meetings.
Python · TypeScript · Node04 / VisionAVSCAFinds objects that break the logic of an aerial scene.
Python · OpenCV · Synthetic dataCurrently
The work gets honest outside.
Last summer I built a drone that carries a ground robot on an NSF research program in Boise. I hold an FAA commercial drone pilot license and flew both Anduril virtual qualifiers. Right now I am focused on making small autonomous aircraft navigate when GPS is unavailable.
Notes from the workbench
Recent writing
On my desk
Reading now
- Mastering Diverse Domains through World ModelsDanijar Hafner, Jurgis Pasukonis, Jimmy Ba, Timothy LillicrapDreamerV3. I want a world model I can actually train a robot inside.
- Autonomous Drone Racing with Deep Reinforcement LearningYunlong Song, Mats Steinweg, Elia Kaufmann, Davide ScaramuzzaRL for racing through gates. Directly useful for Anduril and physical racing.
- Dream-RSI: Recursive Self-Improvement through Evolving WorldsTong Zheng, Xidong Wu, Zheng Zhang, and othersWorld models that keep improving by evolving the worlds they train in.