Notes
Notes on the work behind this site: what was built for the G55, what was measured and what did not work. Newest first. Each says how its numbers were measured and what they cannot show.
- N16's frozen vision tokens do not replace a box detector4 Nov 2026. A head on N16's frozen vision tokens found 2.7 % of vehicles at IoU 0.5, against 55.5 % for YOLO26s on the same frames. Added to the box, the tokens gave a speed error of 3.88 to 4.44 m/s against 2.31 m/s for the analytical tracker.
- Camera position moves where N16 holds the truck in its lane4 Nov 2026. In simulation, moving N16's virtual camera sideways moves its held lane position by -0.97 m per metre of shift. A road-plane warp for a lens 0.239 m right of centre took back about half of the +0.191 m median shift in the hold and restored 0.394 of 1.0 m pushes against 0.197 uncorrected (base 0.509).
- A trained steering correction on N16 recovers more often in simulation4 Nov 2026. On 44 held-out simulator windows, a TD3-trained correction on N16 recovered from 1.5 m pushes 0.76 of the time against 0.58 for its untrained starting point (+0.18 [+0.13, +0.33]). Against the driver's recorded line the gain was +0.07 at 1.5 m and +0.01 at 1.0 m.
- The camera loop ran below 20 Hz because OpenBLAS threads spun3 Nov 2026. On N16 drives the camera loop ran 17.7 to 18.8 Hz: a serial decoder skipped ticks while OpenBLAS worker threads used 1.75 cores. With one OpenBLAS thread per service, three drives held 19.91 to 19.92 Hz.
- comma's models turn the wrong way first because of the base policy31 Oct 2026. In 258 of 258 pushes where a full comma model first steered with the push, its residual-off copy did too. The residual removes 61-88% of these turns but sits at its 1.0 m/s² bound on the rest.
- After a push, N16 still encodes the truck's offset and heading30 Oct 2026. A linear probe reads the truck's lateral offset from N16's vision tokens at R² 0.84-0.93 and from its base summary at 0.87-0.98 on held-out sessions. The wrong-way first turn comes from the base's readout of yaw rate, not from missing information.
- Teaching comma's world model to keep a sideways offset29 Oct 2026. By day comma's world model as shipped already puts the road at a commanded offset (road gain 1.05 at 3 s). Fine-tuning made it overshoot (1.24), reward training did not pass its gate, and no version holds the road at night.
- A truck model learned from the G55's CAN predicts motion better29 Oct 2026. On 12 held-out drives a recurrent network learned from the truck's CAN data has a sideways error of 0.29 m after 4 s of free running, against 7.34 m for the hand-built model and 0.66 m for the same equations refit. It fails a stopping check, so the simulator's default stays hand-built.