Autelica is a San Francisco company founded by Rohan Rao, a robotics engineer who co-wrote the IROS 2021 paper on rebuilding people in 3D from several collaborating drones at Carnegie Mellon’s AirLab. It captures the people doing a job with two to five ordinary cameras, or drones outdoors, measures their motion in 3D with an error grade on every clip, and delivers it ready for a robot, tested in physics. We publish how accurate it is.
Founder
Rohan Rao builds Autelica. At Carnegie Mellon University he took an M.S. in Robotic Systems Development and worked in the AirLab, where he co-wrote 3D Human Reconstruction in the Wild with Collaborative Aerial Cameras (IROS 2021, arXiv 2108.03936). That paper is where Autelica started: several drones filming one person and rebuilding them in 3D.
Before that he co-founded Dynamove, a driver-assistance startup for Indian roads, which placed second at MoveHack 2018 run by NITI Aayog, and studied electrical engineering (B.Tech and M.Tech) at IIT Madras. He has built drones since 2015 and flies under FAA Part 107.
Milestones
- 2021IROS paper on 3D human reconstruction from collaborating aerial cameras, CMU AirLab.
- 31 Jan 2026SkyStudio, the first version of the idea, placed 2nd of 45 teams at the Gemini 3 SuperHack in San Francisco: twelve simulated drones in formation over a stadium, rebuilding players in 3D (SkyStudio demo).
- 30 Sep 2026First three-drone flight round one athlete. Sync, mapping, body fit and rendering all ran on the drones’ footage.
- 1 Oct 2026Motion captured on the drone flight runs on a Unitree G1 in physics: 60 of 60 runs per segment, 59 of them with random pushes and delays.
- 8 Oct 2026A G1 trained on a three-camera capture reaches 12.9 cm keypoint error against 12.3 cm for one trained on a 27-camera mocap dome, in simulation (evaluation-only data; details).
How we measure
Every number we publish names the data, the reference and the method. Public numbers come from ASPset-510, a CC0 dataset of athletes filmed by three fixed 4K cameras with a 3D reference, from our own drone flights, and from robot tests in MuJoCo physics. Every clip we deliver is graded by hiding a camera and checking the reconstruction against what that camera saw. The full list is on the results page.
What we hold to
- Motion is measured from the cameras, never generated, and every clip carries its error.
- Everyone we film consents, and their data is licensed for the use it is put to.
- We don’t call anything live or real-time that isn’t.
- We use data and models whose licences allow commercial use in the product; non-commercial ones only for benchmarks.
Contact
Write to rohan@autelica.com or use the contact form. We are on X and Instagram.