Make your simulator match the real world
Show Tether how things really move. It finds what your simulator gets wrong and hands you a fixed one.
Pick a field
About this field
No video at hand? Simulate your table
Set the physics of the "real" table, and NVIDIA Newton produces the measurements: a phone-style video it renders, or a robot's push log. Tether then has to work out your settings from the data alone; you see how close it got at the end.
Film your own table
- Lay a sheet of A4 or Letter paper flat on the table, beside the path.
- Hold the phone still (or prop it up) so the sheet and the whole path are in view.
- Flick the object 6–10 times from the same spot, short and long. Let it stop each time.
- Use slow motion (120 fps) or 60 fps if your phone has it: motion blur at 30 fps makes launch speeds read low.
Or upload a robot log
CSV or JSON, one row per push. stop in metres and a command (or measured launch_speed) are required; target, perceived and a track (positions at 30 fps, space separated) let it separate actuator, friction and camera errors.
command,stop,target,perceived,track
2.91,0.519,0.520,0.538,0 0.080 0.155 …
Measure
Find what differs
An agent works through your measurements like a scientist: inspect the evidence, propose which physical effects exist, fit each candidate, and ask for more pushes where the data can't decide. The numbers always come from a least-squares fit and a bootstrap over your pushes, never from the language model.
Working
Finding what differs
What we found
Your calibrated simulator
Go deeper
For engineers: how each part of the answer was made.
Every number, with how sure it is
What differs from your simulator
Measured value with its 90% interval (bootstrap: refit on resampled pushes). Grey tick: your current simulator.
Next experiment
Where the plausible models disagree most about where the object stops, inside the range your robot works in.
The evidence: stops, friction along the path, success by distance
Retrain and test
Take it to your simulator
Download a file for your simulator. Each value stays inside the range it was measured in, nothing wider.
1 · Pick where it goes
Preview the file
2 · Check it before you trust it