Measurement engine
Finds each object in every frame and tracks it. From one known reference and the camera geometry, it solves for size, distance, speed and acceleration. Each answer is checked against depth readings and gravity.
The future of robotics will be measured.
Thesis
Robots learn from pixels. They have to act in metres.
A robot that misjudges a step by ten centimetres falls. Most of the data it learns from can't tell it the difference. Video, game footage and simulation describe what things look like. They rarely say how big an object is, how far away it sits or how fast it is moving.
We build data that carries real physical units: object sizes, distances, speeds, accelerations and camera geometry, in metres and seconds. We also build the system that recovers those numbers from ordinary video, so the data can grow beyond what any lab can record by hand.
What we build
Two sources of ground truth, one data format.
Sublevel
A drone mine-rescue simulator. Operators survey collapsed workings, mark hazards and plan rescues. Every decision is logged against the true map, so each flight becomes a labelled episode.
Fly a mission ↗Datasets and evaluation
Episodes, expert preference pairs and benchmark scenarios, exported in the formats robotics teams already train on.
ROS 2 MCAP · LeRobot · RLDS
PCD / LAS point clouds · preference JSONL
PCD / LAS point clouds · preference JSONL
Evidence
Measuring physics from video, on open models.
QuantiPhy, a Stanford benchmark with a NeurIPS 2026 challenge, asks for real sizes, distances, speeds and accelerations from short clips. Our pipeline uses open-weight models only.
Mean relative accuracy on the test set (higher is better). Human and GPT-5.1 figures from the QuantiPhy paper. Our score as of 10 October 2026.
Who it's for
Teams whose robots leave the lab.
Autonomy in hard places
Drone and ground-robot teams working in mines, tunnels and disaster sites, where real failures can't be staged.
Robot foundation models
Labs that need physically grounded episodes with true scale, pose and outcome on every step.
World models
Teams that need action-labelled video, expert preference pairs and scenarios to evaluate against.
Contact
Build it with us.
Partner with us
Robotics, autonomy and world-model teams who need physically correct data, scenarios or evaluation.
Partner with usJoin us
Engineers in perception, simulation and controls who want to teach machines the size of the world.
satyam@afterhumans.ai Join us