What the camera has actually been tested on
Forge films a set and derives numbers from it, bar speed, range of motion, jump height, bar path. Four movements have been checked against real lifts. The rest have not. This page says which is which, because a number you cannot audit is not worth training on.
Right now, the numbers are not accurate.
The camera records and saves your video normally. But right now, the numbers it calculates from that video, velocity, range of motion, power, and similar tracked metrics, are not accurate. We're actively calibrating the system. Don't make training decisions based on these numbers until that's done.
Movement by movement
4 of 12 tracked movements have been validated against real footage.
| Movement | Mode | Tested on real lifts |
|---|---|---|
| Back squatFilmed from behind | Bar path | Yes |
| Pendlay row | Bar path | Yes |
| Bench pressThe hardest case, calibration needs the full body in frame, and a lying athlete often is not | Bar path | Yes |
| Box jump | Jump | Yes |
| DeadliftNext in line | Bar path | Not yet |
| Med ball throwsNext in line | Med ball | Not yet |
| Sprint | Sprint | Not yet |
| Sled push / horizontal load | Horizontal load | Not yet |
| Kettlebell swing | KB swing | Not yet |
| Golf / baseball swing | Rotation | Not yet |
| Running mechanics | Mechanics | Not yet |
| Olympic liftsDeliberately not offered, needs its own path model, see below | None yet | Not yet |
Three things a camera cannot do, whatever the software claims
The angle decides which numbers exist
One camera sees two dimensions. Depth is estimated, and it is the least reliable number in the pipeline. Filming a squat from behind puts forward-and-back drift on exactly that estimated axis, so the reading most coaches want from that angle is the one the setup is worst at. Forge records the angle a set was filmed from and says so next to the number, rather than presenting all three axes as equals.
Olympic lifts need a different model, so we do not offer them
Bar path deviation measures departure from a straight vertical line, and peak velocity is computed on the vertical component alone. That is right for a squat, bench, row or deadlift, where sideways bar movement is error. A correct clean or snatch has a deliberate S-curve. Under a bar-path model, a technically perfect lift reports large deviation, a lift with no curve at all reports clean, and peak velocity understates the bar at exactly the moment that matters. Switching bar-path tracking on for Olympic lifts would produce confident, wrong numbers, so it is off.
Scale comes from the body, and bench is where that breaks
Turning pixels into centimetres needs a known real-world length in frame, and Forge calibrates from the athlete's own height. An athlete lying flat with their feet out of frame cannot produce that measurement. So on bench specifically, a wrong number is more likely to be a calibration failure than a tracking failure, and from outside, the two look identical. A capture that cannot trust its scale is saved with a record of why rather than quietly reported as a result.
What happens when a capture fails
It is kept. A take the tracker could not read is not thrown away, the video is saved for the coach, the set is recorded with whatever could be counted, and the failure is stored with the reason. An empty result and a set nobody filmed are different things, and a system that cannot tell them apart cannot be debugged or improved.