What was asked
N16 is comma's driving model (96244b1f_12800, "N16 nav"). Its vision part turns each pair of camera frames into 32 tokens, a 4 x 8 grid of 512 numbers each. Two questions were written down, with their pass bars, before any job ran.
- Can a small head trained on those frozen tokens alone find vehicles as well as YOLO26s, the box detector the truck runs now?
- Do the tokens, added to YOLO26s's box, give a better speed and heading for a tracked vehicle than the current analytical tracker, a Kalman filter run on the box positions?
What was done
For question 1 the head (a two-layer transformer that outputs a heatmap, offsets and sizes) was trained on labels where at least two of three larger detectors agreed, and scored against 266 hand-checked frames from 7 drives, which held 335 vehicles inside the model views' footprint (508 over the whole image). A vehicle counts as found at an intersection-over-union (IoU) of 0.5 or more. Controls: tokens shuffled in time, and tokens with a box signal planted at 5 %, 1 % and 0.2 % of the token variance, to show whether the head can read a signal at all.
For question 2 the head received N16's tokens at the tracked box over the last 1.6 s. Truth came from each vehicle's whole recorded track, future included, giving 182 moving and 55 parked objects (about 14,500 rows) on 5 drives. Each held-out fold was scored with 3 seeds, and intervals are 90 % bootstrap over drives. Speed error is the primary measure.
Result
Question 1: the tokens alone do not replace YOLO26s.
| Arm | Recall / precision, IoU 0.5 | Recall, IoU 0.2 |
|---|---|---|
| YOLO26s | 55.5 % / 97.6 % | not reported |
| Planted 5 % | 39.4 % / 33.8 % | 71.9 % |
| Planted 0.2 % | 11.6 % / 10.6 % | 52.2 % |
| Real N16 tokens | 2.7 % / 3.5 % | 10.7 % |
| Shuffled tokens | 0.0 % / 0.0 % | 2.4 % |
The head reads planted boxes, so the head is able to see a signal. The real tokens fall below even the 0.2 % plant. The tokens come from a model trained for the driving path, lane lines and one lead car, so individual boxes for parked, oncoming and side-lane vehicles appear to be discarded.
Question 2: the tokens carry some motion. Real tokens had a speed error 1.29 m/s lower than time-shuffled tokens (interval -3.03 to -0.28). The analytical tracker still scored better on every measure:
| Measure | Tracker | Tokens plus query | Tokens plus box history |
|---|---|---|---|
| Speed error (m/s) | 2.31 | 4.44 | 3.88 |
| Moving heading error | 23.0 degrees | 104.2 degrees | 92.9 degrees |
| Parked read as moving | 13.4 % | 45.6 % | 36.4 % |
The third row is the share of parked cars read as moving. A third learned head that uses box history without tokens reached 23.8 %, also worse than the tracker. The analytical tracker therefore stays, with YOLO26s. The recall figures in question 1 count vehicles inside the model views' footprint.
What it does not show
- It does not show that a differently trained head, a finer token grid or a larger model could not find vehicles. One head configuration was run, with no search.
- The truth for question 2 is thin: 5 drives, 2 camera setups, and one drive supplies most rows. Every result is exploratory.
- For parked cars facing nose-in or parallel (39 labelled), the two token heads reached 53.0 % and 51.3 % and the box-history head 49.6 %; the tracker's standing-heading rule reached 59.0 %. With that count, no interval is given.
- Nothing about N16 or YOLO26s was changed, and nothing was run on the truck.