lebowski
2026-07-01 Merged to openpilot master 2026-07-01
Reference model; the G55 drives with it. Highest lateral action rate and shortest lateral lead of the models scored.
Key numbers
The tick is lebowski. Lateral action rate and lateral lead have no better direction; on the others lower is better.
All scores
better, 95% interval excludes zero Better, 95% interval excludes zeroworse, 95% interval excludes zero Worse, interval excludes zerolower, 95% interval excludes zero; no better directionhigher, 95% interval excludes zero; no better direction Lower or higher, no better direction (action rate, lead)not separated, interval includes zero Not separatedno interval (too few drives) No interval: fewer than 10 drives
comma's public drives, open loop
39 segments from 23 drives. Per-segment CSV.
Measured. comma's public two-camera segments (openpilot CI data), native intrinsics and the device's own calibration, both cameras; every model on the same decoded frames. Truth is what the recording car did next: carState speed and acceleration, yaw rate from the device's cameraOdometry. Segments where the car never moves have no data. 39 of the 45 listed segments have a score, from 23 drives and 20 cars.
| Metric | Mean |
|---|---|
| Lateral errorm/s², lower is better | 0.109[0.085, 0.134]median 0.087 |
| Lateral error at the model's own best leadm/s², lower is better | 0.101[0.078, 0.126]median 0.083 |
| Lateral correlation with the car's later motionhigher is better | 0.858[0.793, 0.921]median 0.961 |
| Lateral gain (1 = same size as the car's motion)closer to 1 is better; the delta is the change in distance from 1 | 0.834[0.745, 0.922]median 0.900 |
| Lateral action rate, RMS (smoothness)m/s³, no better direction | 0.664[0.541, 0.792]median 0.644 |
| Lead of the lateral request over the car's motions, no better direction | 0.373[0.316, 0.444]median 0.350n = 35 |
| Longitudinal errorm/s², lower is better | 0.308[0.266, 0.353]median 0.276 |
| Longitudinal correlationhigher is better | 0.718[0.622, 0.803]median 0.814 |
| Longitudinal error while brakingm/s², lower is better | 0.434[0.379, 0.496]median 0.431n = 30 |
| Longitudinal action rate, RMS (smoothness)m/s³, no better direction | 1.33[1.17, 1.51]median 1.31 |
| Plan speed error at 3 sm/s, lower is better | 0.716[0.610, 0.835]median 0.676 |
comma's public drives, closed loop
28 segments from 19 drives. Per-segment CSV.
Measured. comma's worldmodel-4B generates each next frame at 5 Hz from 9 history frames, 5 recorded anchor frames and the car's pose; the model drives comma's ideal car (the requested curvature and acceleration happen at once) for 7.2 s from the recorded start. Distance is to the recording car's own path. One window per segment, nearest the middle, where the car holds at least 5 m/s for the window and the 4.8 s before it; segments with no such stretch are left out. Seeds 0 and 1, averaged per segment. 28 of the 45 listed segments have a score, from 19 drives and 17 cars.
| Metric | Mean |
|---|---|
| Distance from the recorded path at 2 sm, lower is better | 0.148[0.069, 0.281]median 0.072 |
| Distance from the recorded path at 4 sm, lower is better | 0.278[0.125, 0.517]median 0.134 |
| Distance from the recorded path at 7.2 sm, lower is better | 1.04[0.46, 1.78]median 0.31 |
| Share of episodes ever more than 1 m off the pathlower is better | 0.232[0.080, 0.413]median 0.000 |
| Requested lateral jerk, RMSm/s³, lower is better | 0.326[0.228, 0.429]median 0.190 |
| Requested longitudinal jerk, RMSm/s³, lower is better | 0.346[0.289, 0.422]median 0.324 |
| Speed error against the recorded speed, RMSm/s, lower is better | 1.10[0.82, 1.39]median 0.87 |
G55, open loop
4 windows from 3 drives.
Measured. The G55's recorded drives replayed as the truck runs the model: narrow camera only, road-plane warp to 1.45 m (plane:1.45). Truth is the truck's CAN speed and yaw rate. Window = unit. 4 windows: 3 on an earlier camera mount, 1 on the current mount; 2 at night.
| Metric | Mean |
|---|---|
| Lateral errorm/s², lower is better | 0.128no intervalmedian 0.139 |
| Lateral error at the model's own best leadm/s², lower is better | 0.126no intervalmedian 0.138 |
| Lateral correlation with the car's later motionhigher is better | 0.960no intervalmedian 0.963 |
| Lateral gain (1 = same size as the car's motion)closer to 1 is better; the delta is the change in distance from 1 | 0.942no intervalmedian 0.930 |
| Lateral action rate, RMS (smoothness)m/s³, no better direction | 0.957no intervalmedian 0.935 |
| Lead of the lateral request over the car's motions, no better direction | 0.338no intervalmedian 0.350 |
| Longitudinal errorm/s², lower is better | 0.315no intervalmedian 0.309 |
| Longitudinal correlationhigher is better | 0.582no intervalmedian 0.636 |
| Longitudinal error while brakingm/s², lower is better | 0.515no intervalmedian 0.516 |
| Longitudinal action rate, RMS (smoothness)m/s³, no better direction | 1.75no intervalmedian 1.62 |
| Plan speed error at 3 sm/s, lower is better | 1.03no intervalmedian 0.96 |
G55, closed loop
18 windows from 4 drives.
Measured. worldmodel-4B as in the public closed loop, but the model drives a simulation of the G55's measured steering and powertrain response instead of the ideal car. Lane-following windows from the G55's drives, day and night; distance is to the driver's recorded path; seeds 0 and 1, averaged per window. 18 windows (9 day, 9 night).
| Metric | Mean |
|---|---|
| Distance from the recorded path at 2 sm, lower is better | 0.268no intervalmedian 0.227 |
| Distance from the recorded path at 4 sm, lower is better | 0.562no intervalmedian 0.367 |
| Distance from the recorded path at 7.2 sm, lower is better | 1.60no intervalmedian 0.68 |
| Share of episodes ever more than 1 m off the pathlower is better | 0.361no intervalmedian 0.000 |
| Requested lateral jerk, RMSm/s³, lower is better | 0.545no intervalmedian 0.428 |
| Requested longitudinal jerk, RMSm/s³, lower is better | 0.361no intervalmedian 0.340 |
| Speed error against the recorded speed, RMSm/s, lower is better | 1.06no intervalmedian 1.06 |
G55, recovery tests
18 windows from 4 drives.
Measured. The truck closed loop with a disturbance while the model drives; only the tests listed ran. 2° off heading: an extra 2°/s yaw for the first second. 18 windows (9 day, 9 night). 0.5 m off centre: an extra yaw-rate pulse, out and back over the first 2 s, that would move an uncorrected car 0.5 m sideways with no net heading change. 18 windows (9 day, 9 night).
| Metric | Mean |
|---|---|
| 2° off heading: distance from the path at 2 sm, lower is better | 0.427no intervalmedian 0.397 |
| 2° off heading: distance from the path at 4 sm, lower is better | 0.544no intervalmedian 0.366 |
| 2° off heading: requested lateral jerk, RMSm/s³, lower is better | 0.827no intervalmedian 0.806 |
| 0.5 m off centre: distance from the path at 2 sm, lower is better | 0.349no intervalmedian 0.292 |
| 0.5 m off centre: distance from the path at 4 sm, lower is better | 0.659no intervalmedian 0.383 |
| 0.5 m off centre: requested lateral jerk, RMSm/s³, lower is better | 1.17no intervalmedian 1.09 |
Limits
- Open-loop scores compare a model's action with what the recording driver or car did; a model that drives differently but well is scored as wrong. comma's RL residual makes the lateral action smoother, and earlier than its base's by 0.10 to 0.26 s (measured on the G55 and on 16 of comma's cars); both score as error here. Measured on the 2026-10-27 deep dive: open-loop accuracy did not predict closed-loop driving.
- Closed-loop numbers come from a generative world model, not a car. In a separate test on the G55's windows, worldmodel-4B followed an added yaw command through 2 s but showed 0 to 20% of an added lateral offset 3 to 4 s later; past about 4 s the recorded anchor frames also pull every model back toward the recorded path. The front-page ranking uses 2 s.
- Intervals are 95% percentile bootstraps that resample whole drives (the drive a public segment was cut from, the recorded drive a truck window comes from), so segments of one drive count once. Below 10 drives there is no interval: the truck sections come from fewer drives and report means and win counts only.
- Paired deltas are model minus baseline on the units both have (for lateral gain, the change in its distance from 1); wins count units where the model is better in the metric's direction.
- The lead of the lateral request is the shift, searched from 0 to 1.5 s, at which the request best correlates with the car's later lateral motion; a larger value means the request comes earlier. A value at either edge is censored and left out. Plan speed error at 3 s reads the plan head, which does not separate models built on one frozen base: it is identical on every segment for the four f22e06e5 models, and for Cinque v2 and v3.
- Truck open loop pools few windows (camera mounts changed between drives); read it as a check that a model works on the truck's camera, not a ranking.
- Truck closed loop: the steering and powertrain simulation is fitted to the G55, not validated on it in closed loop; the reference is the driver's own line, not the lane centre.
- Recovery tests are disturbances rendered by the world model (yaw it renders; a one-frame jump it does not, so none is used). The braking-lead test is not built yet: the drives the truck closed loop can use hold no window where a car ahead brakes.
The model file
- Model id
lebowski_877m- SHA-256
a501760a9d1d5fef0eab2b8c5d122d06124fc26dc8e0782e0aa94b82a208f0ff- File scored
- openpilot-master,
fa0c6876d3cf070e91e25e5353ceadc68a5b3285(link), 2026-07-01 - First public
- 2026-07-01
- Parameters
- 876.5M
- GMAC per step
- 99.4
- Family
- lebowski
- Scored
- 2026-10-27 21:27 UTC, suite
geval-1, code102710f59c - Notes
- 877M ConvNeXt encoder, host-kept history, no RL residual. Checkpoint 6f5c8d09/ecc73f48. The model the G55 runs.
Results file (results/lebowski_877m.json, schema 1).