BADAS-3.0: Real-Time Driving Perception and Planning from One Camera
Organizations: Empiric Earth
Abstract
We ask how far a driving model can get with one camera and real-time latency. BADAS-3.0 is a 1B-parameter video encoder with four small heads. From two seconds of one forward camera, it predicts the current speed, the metric path just driven, a collision-risk score and six candidate futures over 6.4 seconds, in under 50 ms. Given the car's current speed, BADAS-3.0 has 44% lower minADE over five trajectory benchmarks than the 34B Alpamayo 2 Super, which reads six cameras and runs over 100x slower. The gain comes from camera rigs that neither model was trained on. On the rig they share, the two are on par. Existing benchmarks are mostly nominal driving, so we add two open-loop safety benchmarks, one of real dashcam collisions and near-misses and one of Euro NCAP scenarios. On both, BADAS-3.0 has a higher pass rate than every published planner we tested.
Figures & tables
| model | input | params | nuSc | AV2-S | AV2-TbV | Waymo | PAI-644 | mean 5 | top-1 |
|---|---|---|---|---|---|---|---|---|---|
| BADAS-3.0 , speed head, zero-shot | 1 cam | 1B | 1.223 | 1.440 | 1.286 | 3.192 | 1.272 | 1.682 | 3.912 |
| BADAS-3.0 , speed head, few-shot | 1 cam | 1B | 1.217 | 1.234 | 1.061 | 1.518 | 1.275 | 1.261 | 3.300 |
| BADAS-3.0 , speed sensor | 1 cam | 1B | 0.896 | 1.006 | 0.942 | 1.000 | 0.929 | 0.955 | 2.719 |
| BADAS-3.0 , speed sensor + ego history | 1 cam | 1B | 0.894 | 1.005 | 0.946 | 0.988 | 0.927 | 0.952 | 2.711 |
| Alpamayo 2 Super [ 48 ] | 6 cam | 34B | 2.347 | 1.977 | 1.774 | 1.414 | 0.957 | 1.694 | 3.423 |
| Alpamayo-R1 [ 44 ] | 4 cam | 10B | 2.138 | 2.159 | 2.075 | 1.692 | 1.031 | 1.819 | 3.741 |
| method | scale | nuSc | AV2-S | AV2-TbV | Waymo | KITTI | PAI-644 | mean 6 | runtime |
|---|---|---|---|---|---|---|---|---|---|
| BADAS-3.0 , speed head, zero-shot | metric | 0.44 | 0.99 | 0.82 | 1.53 | 3.03 | 0.25 | 1.18 | 43 ms |
| BADAS-3.0 , speed head, few-shot | metric | 0.43 | 0.40 | 0.31 | 0.47 | 0.53 | 0.25 ‡ | 0.40 ‡ | 43 ms |
| XVO [ 27 ] | metric | 0.81 | 2.62 | 1.95 | 1.72 | 2.02 | 2.22 | 1.89 | 190 ms |
| ZeroVO [ 28 ] | metric | 2.36 | 1.52 | 1.21 | 2.25 | 2.98 | 4.53 | 2.47 | 9.4 s |
| OpenVO [ 43 ] | metric | 3.76 | 3.40 | 1.88 | 2.05 | 3.56 | 3.60 | 3.04 | 5.8 s |
| UniDepth-V2 + PnP [ 53 ] | metric | 3.97 | 2.04 | 1.61 | 0.86 | 1.19 | 2.13 | 1.97 | 1.0 s |
| model | params | chal. 1w | chal. slide | long-tail | DAD | DoTA | DADA |
|---|---|---|---|---|---|---|---|
| =1344 | =1344 | =884 | =314 | =367 | =113 | ||
| BADAS-3.0 risk head | 1B | 0.949 | 0.892 | 0.995 | 0.916 | 0.999 | 0.996 |
| BADAS-2.0 [ 22 ] | 300M | 0.940 | 0.876 | 0.994 | 0.922 | 0.999 | 0.996 |
| BADAS-2.0-Flash | 86M | 0.941 | 0.879 | 0.990 | 0.849 | 0.998 | 0.995 |
| BADAS-2.0-Flash-Lite | 22M | 0.933 | 0.867 | 0.984 | 0.870 | 0.998 | 0.992 |
| BADAS-1.0 [ 21 ] | 300M | 0.925 | 0.861 | 0.960 | 0.885 | 0.998 | 0.989 |
| Euro NCAP | DCA-Bench | |||||
| model | cameras | score | FP rate (%) | pass rate (%) | FP rate (%) | pass rate (%) |
| Full brake, 6 m/s 2 (control) | – | 1.000 | 100.0 | 0.0 | 93.3 | 6.7 |
| Recorded driver | – | – | – | – | 0.0 | 44.9 |
| BADAS-3.0 | 3 front, 150 ∘ | 0.365 | 1.4 | 19.1 | – | – |
| BADAS-3.0 | 1 front, 75 ∘ | 0.296 | 0.0 | 15.4 | 13.5 | 39.3 |
| Alpamayo 2 Super † | 6 | 0.202 | 0.1 | 9.8 | 13.1 | 14.2 |
| Euro NCAP | DCA-Bench | |||||
|---|---|---|---|---|---|---|
| BADAS-3.0 | cameras | score | FP rate (%) | pass rate (%) | FP rate (%) | pass rate (%) |
| – | 1 front, 75 ∘ | 0.193 | 0.0 | 4.8 | 13.5 | 20.2 |
| with selector fine-tune | 1 front, 75 ∘ | 0.296 | 0.0 | 15.4 | 13.5 | 39.3 |
| – | 3 front, 150 ∘ | 0.224 | 0.0 | 6.7 | – | – |
| with selector fine-tune | 3 front, 150 ∘ | 0.365 | 1.4 | 19.1 | – | – |
Appendix figures & tables25 assets
Supplementary material from the paper’s appendix.
Appendix
| method | scale | nuSc | AV2-S | AV2-TbV | Waymo | KITTI | PAI-644 | mean 6 | runtime |
|---|---|---|---|---|---|---|---|---|---|
| BADAS-3.0 , speed head, zero-shot | metric | 0.033 | 0.031 | 0.023 | 0.040 | 0.057 | 0.024 | 0.035 | 43 ms |
| BADAS-3.0 , speed head, few-shot | metric | 0.034 | 0.032 | 0.023 | 0.041 | 0.057 | 0.025 ‡ | 0.035 ‡ | 43 ms |
| XVO [ 27 ] | metric | 0.063 | 0.102 | 0.108 | 0.063 | 0.118 | 0.253 | 0.118 | 190 ms |
| ZeroVO [ 28 ] | metric | 0.074 | 0.132 | 0.217 | 0.056 | 0.118 | 0.479 | 0.179 | 9.4 s |
| OpenVO [ 43 ] | metric | 0.060 | 0.091 | 0.123 | 0.038 | 0.076 | 0.478 | 0.145 | 5.8 s |
| UniDepth-V2 + PnP [ 53 ] | metric | 1.685 | 0.258 | 0.392 | 0.256 | 0.508 | 0.397 | 0.583 | 1.0 s |
| method | nuSc | AV2-S | AV2-TbV | Waymo | KITTI | PAI-100c | mean 6 |
|---|---|---|---|---|---|---|---|
| BADAS-3.0 speed head, zero-shot | 0.50 | 0.84 | 0.75 | 1.68 | 3.22 | 0.55 | 1.26 |
| BADAS-3.0 speed head, few-shot | 0.50 | 0.46 | 0.41 | 0.70 | 0.56 | 0.52 | 0.52 |
| XVO | 1.02 | 2.14 | 1.95 | 2.31 | 2.32 | 4.36 | 2.35 |
| ZeroVO | 1.94 | 2.76 | 2.72 | 2.53 | 3.27 | 6.67 | 3.31 |
| OpenVO | 3.52 | 7.05 | 4.87 | 3.51 | 3.64 | 4.99 | 4.60 |
| UniDepth-V2 + PnP | 17.09 | 5.93 | 4.39 | 65.90 | 3.82 | 7.15 | 17.38 |
| no alignment | similarity fit | |||||||
| method | PAI-644 | AV2 | Waymo | Prop. | PAI-644 | AV2 | Waymo | Prop. |
| BADAS-3.0 , stitched, zero-shot | 1.82 | 2.44 | 10.07 | 4.94 | 0.33 | 0.54 | 0.52 | 0.91 |
| BADAS-3.0 , stitched, few-shot | 1.62 ‡ | 2.37 | 4.70 | 10.65 ‡ | 0.34 ‡ | 0.54 | 0.68 | 0.94 ‡ |
| BADAS-3.0 , chained, zero-shot | 2.11 | 2.82 | 12.10 | 6.03 | 0.42 | 0.66 | 0.76 | 1.15 |
| BADAS-3.0 , chained, few-shot | 1.96 ‡ | 2.73 | 4.15 | 11.80 ‡ | 0.44 ‡ | 0.64 | 0.91 | 1.12 ‡ |
| BADAS-3.0 , stitched, speed sensor (ref.) | 1.03 | 1.27 | 2.53 | 1.98 | 0.22 | 0.36 | 0.44 | 0.66 |
| model | nuSc | AV2-S | AV2-TbV | Waymo | PAI-644 | mean 5 |
|---|---|---|---|---|---|---|
| BADAS-3.0 , speed head, zero-shot | 1.22 0.12 | 1.44 0.14 | 1.29 0.10 | 3.19 0.40 | 1.27 0.14 | 1.682 |
| BADAS-3.0 , speed head, few-shot | 1.22 0.12 | 1.23 0.15 | 1.06 0.09 | 1.52 0.20 | 1.27 0.14 | 1.261 |
| BADAS-3.0 , speed sensor | 0.90 0.09 | 1.01 0.13 | 0.94 0.08 | 1.00 0.11 | 0.93 0.06 | 0.955 |
| BADAS-3.0 , speed sensor + ego history | 0.89 0.09 | 1.00 0.13 | 0.95 0.09 | 0.99 0.11 | 0.93 0.06 | 0.952 |
| Alpamayo 2 Super | 2.35 0.30 | 1.98 0.34 | 1.77 0.20 | 1.41 0.24 | 0.96 0.09 | 1.694 |
| Alpamayo-R1 | 2.14 0.22 | 2.16 0.41 | 2.08 0.22 | 1.69 0.23 | 1.03 0.09 | 1.819 |
| slot (index) | rig heading | nuScenes | AV2 Sensor and TbV | Waymo | PAI-644 |
|---|---|---|---|---|---|
| cross_left (0) | 67 ∘ | CAM_FRONT_LEFT (55 ∘ ) | ring_front_left (45 ∘ ) | FRONT_LEFT (45 ∘ ) | cross_left |
| front_wide (1) | 0 ∘ | CAM_FRONT | ring_front_center | FRONT | front_wide |
| cross_right (2) | CAM_FRONT_RIGHT ( ) | ring_front_right ( ) | FRONT_RIGHT ( ) | cross_right | |
| rear_left (3) | 152 ∘ | CAM_BACK_LEFT (109 ∘ ) | ring_rear_left (153 ∘ ) | SIDE_LEFT (90 ∘ ) | rear_left |
| rear_right (5) | CAM_BACK_RIGHT ( ) | ring_rear_right ( ) | SIDE_RIGHT ( ) | rear_right | |
| front_tele (6) | 0 ∘ , 30 ∘ FOV | central crop of CAM_FRONT | central crop of ring_front_center | central 30 ∘ crop of FRONT | front_tele |
| model | nuSc | AV2-S | AV2-TbV | Waymo | PAI-644 |
|---|---|---|---|---|---|
| constant velocity | 3.714 | 4.094 | 3.800 | 3.497 | 6.674 |
| BADAS-3.0 , head’s own pick | 2.610 | 2.794 | 2.714 | 2.713 | 2.726 |
| Alpamayo 2 Super | 5.036 | 4.498 | 3.923 | 2.723 | 2.288 |
| Alpamayo-R1 | 4.777 | 4.922 | 4.152 | 3.108 | 2.301 |
| Alpamayo 1.5 | 4.395 | 4.374 | 3.867 | 2.543 | 2.075 |
| set | run | minADE 6 | first sample | mean of six |
|---|---|---|---|---|
| nuSc | Alpamayo 1.5, seed 42 | 2.117 | 4.395 | 4.292 |
| nuSc | Alpamayo 1.5, seed 43 | 1.859 | 4.677 | 4.458 |
| nuSc | Alpamayo 1.5, seed 44 | 1.924 | 3.687 | 4.097 |
| nuSc | Alpamayo 1.5, three seeds | 1.967 0.134 | 4.253 0.510 | 4.282 0.181 |
| PAI-644 | Alpamayo 1.5, seed 42 | 0.954 | 2.075 | 1.987 |
| PAI-644 | Alpamayo 1.5, seed 43 | 0.921 | 2.037 | 1.991 |
| model | input | params | nuSc | AV2-S | AV2-TbV | Waymo | PAI-644 | mean 5 |
| BADAS-3.0 , speed head, zero-shot | 1 cam | 1B | 0.928 | 1.114 | 1.023 | 2.690 | 0.983 | 1.347 |
| BADAS-3.0 , speed head, few-shot | 1 cam | 1B | 0.916 | 0.939 | 0.798 | 1.221 | 0.983 | 0.971 |
| BADAS-3.0 , speed sensor | 1 cam | 1B | 0.610 | 0.653 | 0.637 | 0.672 | 0.637 | 0.642 |
| BADAS-3.0 , speed sensor + ego history | 1 cam | 1B | 0.612 | 0.654 | 0.643 | 0.661 | 0.631 | 0.640 |
| Alpamayo 2 Super [ 48 ] | 6 cam | 34B | 1.582 | 1.305 | 1.179 | 0.923 | 0.623 | 1.122 |
| Alpamayo-R1 [ 44 ] | 4 cam | 10B | 1.509 | 1.455 | 1.420 | 1.131 | 0.671 | 1.237 |
| model | cameras | score | pairs passed | FP rate (%) | |||||||
| s | s | s | mean | s | s | s | s | s | s | ||
| Full brake, 6 m/s 2 (control) | – | 1.000 | 1.000 | 1.000 | 1.000 | 0 | 0 | 0 | 100.0 | 100.0 | 100.0 |
| BADAS-3.0 | 3 front, 150 ∘ | 0.401 | 0.423 | 0.272 | 0.365 | 143 | 131 | 68 | 2.5 | 2.5 | 1.4 |
| BADAS-3.0 | 1 front, 75 ∘ | 0.357 | 0.317 | 0.214 | 0.296 | 135 | 115 | 55 | 0.0 | 0.0 | 0.0 |
| BADAS-3.0, head’s own pick | 3 front, 150 ∘ | 0.296 | 0.229 | 0.146 | 0.224 | 111 | 72 | 24 | 0.3 | 0.0 | 0.0 |
| Alpamayo 2 Super † | 6 | 0.220 | 0.222 | 0.164 | 0.202 | 85.7 | 71.7 | 35.0 | 0.5 | 0.2 | 0.1 |
| metric | BADAS-3.0 | without floor | with floor (Eq. 1 ) |
|---|---|---|---|
| Table 1 mean 5 (m) | speed head, zero-shot | 2.545 | 2.532 |
| Table 1 mean 5 (m) | speed head, few-shot | 1.490 | 1.476 |
| Table 1 mean 5 (m) | speed sensor | 1.120 | 1.108 |
| Table 1 mean 5 (m) | speed sensor + ego history | 1.111 | 1.098 |
| Table 1 top-1 (m) | speed head, zero-shot | 5.078 | 5.064 |
| Table 1 top-1 (m) | speed head, few-shot | 3.600 | 3.573 |
| incidents passed (%) | twins | |||
| model | published | drivable | FP rate (%) | pairs passed (%) |
| Full brake, 6 m/s 2 (control) | 98.0 | 98.0 | 93.3 | 6.7 |
| Recorded driver | 45.0 | 45.0 | 0.0 | 44.9 |
| BADAS-3.0 | 56.0 | 46.0 | 13.5 | 39.3 |
| BADAS-3.0, head’s own pick | 32.0 | 25.0 | 13.5 | 20.2 |
| Alpamayo-R1 † | 19.5 | 19.5 | 9.7 | 18.2 |
| backbone | params | frozen, last layer | frozen, four depths | fine-tuned | risk AP | speed MAE | ego-pose ATE |
|---|---|---|---|---|---|---|---|
| Driving V-JEPA ViT-g | 1.01B | 2.198 | 2.047 | 1.499 | 0.893 | 1.27 | 0.235 |
| V-JEPA 2.1 ViT-G | 1.90B | 2.273 | 2.352 | 1.735 | 0.848 | 1.58 | 0.281 |
| V-JEPA 2 ViT-g | 1.01B | 2.029 | 2.372 | 1.750 | 0.877 | 1.89 | 0.257 |
| V-JEPA 2.1 ViT-g | 1.01B | 2.522 | 2.396 | 1.865 | 0.853 | 2.08 | 0.278 |
| Cosmos-Embed1 | 0.99B | 2.355 | 2.447 | 2.224 | 0.490 | 2.72 | 0.363 |
| model | params | input | GPU | median latency | peak memory |
|---|---|---|---|---|---|
| BADAS-3.0 , compiled | 1B | 1 cam | A100 | 43 ms | 1.9 GiB |
| BADAS-3.0 , eager | 1B | 1 cam | A100 | 146 ms | 2.0 GiB |
| Qwen-Drive 4B, direct | 4B | 3 cam | A100 | 967 ms | 12.3 GiB |
| Qwen-Drive 4B, reasoning | 4B | 3 cam | A100 | 2.0 s | 11.2 GiB |
| Alpamayo-R1 | 10B | 4 cam | A100 | 4.1 s | 25.9 GiB |
| Alpamayo 1.5 | 10.5B | 4 cam | A100 | 4.4 s | 26.1 GiB |
| method | sensors | PDMS |
|---|---|---|
| no reinforcement learning | ||
| TransFuser [ 10 ] | 3 cam + LiDAR | 84.0 |
| Epona [ 70 ] , world model | 1 cam | 86.2 |
| DiffusionDrive [ 33 ] | 3 cam + LiDAR | 88.1 |
| Qwen-Drive-1.0 [ 73 ] , SFT, reasoning | 3 cam | 88.2 |
| BADAS-3.0, recalibrated | 1 cam | 89.0 |
| method | best-of- PDMS | |
|---|---|---|
| Qwen-Drive-1.0 [ 73 ] , SFT | 6 | 89.3 |
| Qwen-Drive-1.0 [ 73 ] , RL | 6 | 91.4 |
| AutoVLA [ 74 ] | 6 | 92.1 |
| DriveVLA-W0 [ 31 ] | 6 | 93.0 |
| AdaThinkDrive [ 39 ] | 4 | 93.0 |
| ExploreVLA [ 56 ] | 6 | 93.7 |
| held-out cell | zero-shot | earlier rule | frozen rule | ( , se) | result | with the 5% minimum | |
|---|---|---|---|---|---|---|---|
| KITTI, 2 s | 102 | 3.059 | 0.525 | 0.538 | (1.491, 0.041) | pass | 0.538 (unchanged) |
| PAI-100c, 2 s | 187 | 0.312 | 0.327 | 0.367 | (1.062, 0.012) | fail, | 0.367 (still applied) |
| PAI-100c, 15 s stitched | 90 | 2.361 | 2.501 | 2.504 | (1.046, 0.014) | fail, | zero-shot (gate closed) |
| PAI-100c, 15 s chained | 90 | 2.656 | 2.836 | 2.582 | (1.052, 0.015) | pass, | 2.582 (still applied) |
Autonomous Driving
Autonomous driving develops vehicles that perceive road scenes, predict interactions, and plan and control safe movement with limited or no human input.
See all 576 papers