Anticipates car crashes. A risk score that ramps as the danger develops.
A 22M-parameter collision-anticipation layer. Give it an ego scene and it returns how likely a collision is over the seconds ahead, scored continuously, plus a map of what the model was looking at when it decided. Production-grade for AV programs, fleets, and ADAS integrations.
22M params<3 ms on A100100M miles / month350,000 edge devicesZero synthetic data
Fleets & Logistics
Trucks, taxis, automobiles
Legged
Quadrupeds
Industrial
Forklifts
Sidewalk
Sidewalk delivery
Defence
Off-road defence
01 – Live Demo
Drop your video. See collision probability per frame.
Upload a clip from any camera that moves through the physical world – vehicle, forklift, sidewalk delivery, drone. BADAS 2.0 outputs a graded probability per frame. Your planner decides what to do with it.
Real anonymized edge device incidents – BADAS 2.0 prediction overlay on actual near-collision events from Empiric Earth's 350K camera fleet
BADAS 2.0 collision probability overlay – probability ramps from near-zero as danger develops. Click to play.
Generational leap – BADAS 2.0 detects danger earlier with higher confidence and 74% fewer false positives
BADAS 2.0 vs COSMOS-BADAS – purpose-built vision transformer outperforms fine-tuned VLM on temporal discrimination
Attention heatmap – visualizes where the model focuses as danger develops
BADAS-Reason – natural language explanations describe what the model sees and why it predicts danger. Tap to play with sound.
02 – Beyond the Road
A layer, not a stack. Drop it onto the machine.
BADAS 2.0 outputs a graded collision probability per frame. Your planner decides what to do with it. We don't replace the certified stack – we sit beside it as the long-tail anticipation layer.
Runs zero-shot on machines it has never seen.
Zero-shot, with no retraining. The same 22M-parameter Flash Lite we ship for ADAS lights up conflicts on platforms it was never trained on.
22M
params, zero-shot
Zero-shot
Quadrupeds, forklifts, sidewalk POV, off-road defence – environments never in training. No retraining required.
<3 ms
Flash Lite on NVIDIA A100; 5.9 ms on Jetson Thor. 12× faster than the 300M reference at near-parity AP.
100M mi/mo
Trained on 100M real-world miles a month from 350,000 active edge devices. No synthetic data.
Tell us what your machine sees.
Drop a clip and book a 30-minute scoping call. We'll show you the per-frame probability on your environment.
The headline numbers, in priority order. Per-category benchmark tables below.
01
22M params. <3 ms on A100. 5.9 ms on Jetson Thor.
Scored where the sensor is. No connectivity, no round trip, and it lands on the silicon your machine already shipped on.
02
A graded collision probability, scored continuously.
It ramps as the danger develops, which is what makes it a warning rather than a report.
03
Generalizes across machines, with no retraining.
Quadrupeds, forklifts, sidewalk POV, off-road defence vehicles. Same model. Same weights.
04
A layer, not a stack.
We don't replace the certified stack. Your planner decides what to do with the score.
05
100M real-world miles a month, 350,000 active edge devices. Zero synthetic.
Trained on real long-tail conflict, not generated frames.
Detailed benchmarks
Long-Tail Benchmark: Per-Category Breakdown
AUC and AP across 10 scenario groups (888 clips, sliding window). BADAS 2.0 leads in every category. Best per row in bold.
Group
BADAS 2.0
BADAS 2.0Flash
BADAS 2.0Flash Lite
BADAS 1.0
BADAS Open
COSMOS-Reason2Fine-tuned
Gemini 2.5 ProFine-tuned
Qwen3-VL-2B
Animal
AUC
96.4%
94.8%
92.3%
94.8%
88.1%
81.6%
79.9%
75.9%
AP
99.1%
98.8%
98.1%
98.7%
95.7%
95.2%
93.9%
93.2%
Pedestrian
AUC
99.8%
99.6%
99.1%
99.1%
84.4%
93.6%
77.4%
67.4%
AP
99.9%
99.7%
99.4%
99.4%
87.9%
95.6%
77.5%
75.5%
Intersection
AUC
100.0%
100.0%
99.8%
98.8%
95.8%
97.9%
85.0%
61.7%
AP
100.0%
100.0%
99.7%
98.4%
93.7%
96.9%
73.7%
45.3%
Overtaking
AUC
100.0%
100.0%
99.7%
97.4%
92.7%
97.5%
83.2%
82.6%
AP
100.0%
99.9%
99.4%
96.1%
85.3%
94.8%
60.8%
68.0%
Snow
AUC
100.0%
100.0%
99.9%
99.6%
93.2%
97.5%
95.3%
80.4%
AP
100.0%
100.0%
99.9%
99.5%
94.0%
97.3%
93.6%
77.9%
Infrastructure
AUC
100.0%
99.4%
98.4%
86.9%
84.0%
97.4%
90.8%
58.4%
AP
100.0%
99.6%
98.7%
91.5%
88.0%
98.1%
92.2%
62.5%
Motorcyclist
AUC
99.8%
99.8%
100.0%
99.8%
96.6%
98.9%
80.1%
72.5%
AP
99.9%
99.9%
100.0%
99.9%
96.9%
99.1%
76.6%
76.1%
Cyclist
AUC
100.0%
98.7%
99.4%
98.6%
93.1%
94.0%
82.9%
64.8%
AP
100.0%
99.1%
99.5%
98.7%
94.2%
95.3%
81.2%
59.8%
Rain
AUC
100.0%
100.0%
100.0%
97.5%
96.6%
99.6%
82.2%
81.5%
AP
100.0%
100.0%
100.0%
98.0%
95.9%
99.4%
64.5%
66.2%
Fog
AUC
100.0%
99.8%
99.8%
100.0%
99.4%
98.2%
81.6%
82.5%
AP
100.0%
99.5%
99.5%
100.0%
98.7%
95.9%
60.5%
76.0%
OVERALL
AUC
99.3%
98.9%
98.1%
94.9%
82.3%
92.6%
83.3%
67.8%
AP
99.4%
99.0%
98.4%
96.0%
84.5%
94.1%
79.5%
67.2%
Nexar Kaggle Benchmark
Single-window mean AP over three lead-time thresholds (1,344 clips). BADAS 2.0 improves mAP from 92.5% to 94.0% while cutting the false positive rate by 74%.
Model
AP @0.5s
AP @1.0s
AP @1.5s
mAP
FPR
Params
BADAS 2.0
94.3%
95.7%
92.1%
94.0%
4.6%
300M
BADAS 2.0 Flash
94.5%
96.2%
91.5%
94.1%
9.7%
86M
BADAS 2.0 Flash Lite
94.6%
94.7%
90.7%
93.3%
12.2%
22M
BADAS 1.0
93.5%
93.6%
90.4%
92.5%
10.9%
300M
COSMOS-BADAS
90.4%
88.9%
87.5%
88.9%
–
2B
Reading this table: AP @0.5s / @1.0s / @1.5s – Average Precision measured at three different lead times before the collision. Higher = better detection at that warning horizon. mAP – Mean Average Precision, the average of the three AP scores above. FPR – False Positive Rate. Lower = fewer false alarms. Params – Model size in parameters.
74% Fewer False Alarms
On the internal test set, BADAS 2.0 cuts the false positive rate from 17.7% (v1.0) to 4.6% – a 74% reduction with no loss of recall.
4.6%
FPR – BADAS 2.0
mAP 94.0% · 300M params · 34ms
9.7%
FPR – BADAS 2.0 Flash
mAP 94.1% · 86M params · 4.8ms
10.9%
FPR – BADAS 1.0
mAP 92.5% · 300M params · 2,500ms
Early Warning Recall (Long-Tail Benchmark)
Fraction of collision clips that got an alert before the event (888 clips, 10 scenario groups, threshold 0.75). The F1 beside each bar is a separate measurement on the same run, not derived from the bar: it is the harmonic mean of per-window precision and recall at that threshold, while the bar is measured per clip. F1 counts missed events and false alarms against each other, so a model cannot score well by alerting constantly. Figures from Table 6 of Beyond the Beep: Scalable Collision Anticipation and Real-Time Explainability with BADAS-2.0.
BADAS 2.0
91.3%
F1 96.4%
BADAS 2.0 Flash
89.9%
F1 93.8%
BADAS 1.0
85.5%
F1 87.5%
External Benchmarks (Sliding Window)
AUC and AP on three public academic benchmarks using ego-centric re-annotation. Best per column in bold.
Model
DAD AUC
DAD AP
DoTA AUC
DoTA AP
DADA AUC
DADA AP
BADAS 2.0
99.3%
92.2%
99.1%
99.9%
99.1%
99.6%
BADAS 2.0 Flash
98.7%
84.9%
98.5%
99.8%
99.0%
99.5%
BADAS 2.0 Flash Lite
98.2%
87.0%
98.5%
99.8%
98.1%
99.2%
BADAS 1.0
99.0%
94.0%
72.0%
95.0%
87.0%
90.0%
COSMOS-BADAS
94.4%
60.2%
98.3%
99.8%
95.9%
97.8%
Qwen3-VL-2B
75.4%
14.1%
70.9%
95.1%
80.5%
88.6%
Reading this table: AUC – Area Under the ROC Curve. Measures how well the model separates collisions from safe driving. 100% = perfect. AP – Average Precision. DAD, DoTA, DADA-2000 – Three public academic collision anticipation benchmarks with re-annotated ego-centric protocol.
How Confidence Evolves Over Time
Average collision probability over normalized pre-event time (0% = start, 100% = event). Each clip's timeline is scaled independently, so clips of different lengths are comparable. Positive clips only. BADAS models ramp up sharply; competitors stay flat.
Reading this chart: For every positive clip, each model's prediction timeline is normalized so 0% = first prediction and 100% = labeled event. Predictions are binned into 10 equal intervals then averaged across clips. Curves are baseline-normalized per model so the y-axis shows each model's rise above its own floor. A steep ramp means confidence increases sharply as the event approaches; a flat line means the model outputs a near-constant score regardless of proximity to collision.
04 – See What the Model Sees
A graded probability the planner can act on.
BADAS 2.0 outputs an attention heatmap and a per-frame collision probability. Drop both into your planner, your fleet console, or your audit log. The model is a layer; the planner decides.
Where the model is looking
Attention heatmap on a sidewalk-delivery POV clip – the focus shifts to the pedestrian crossing the robot's path before the conflict materializes. Same overlay you saw on a forklift in section 1.
Sidewalk-delivery POV – attention shifts to the pedestrian before the conflict
Why it called the conflict
BADAS-Reason explains the call in natural language. For a sidewalk-delivery POV like the clip on the left, the description names the actor (pedestrian), direction, and predicted intersection – useful for fleet review and audit, not a control signal.
BADAS-Reason – natural language description of the predicted conflict
05 – Silicon
Runs on the silicon you've already shipped.
Three sizes of the same model, validated on the silicon robotics teams actually deploy on – A100 in the cloud and Jetson Thor on the edge. Same architecture, same training, same world model.
BADAS 2.0
300M parameters
Highest accuracy across the long tail
Best mTTA and early warning recall
For cloud analytics and offline scoring
34 ms A100 · 41 ms Jetson Thor
BADAS 2.0 Flash
86M parameters
Rack-edge model for robotics integrators
Outperforms BADAS 1.0 on every metric
4.8 ms A100 · 12.5 ms Thor
BADAS 2.0 Flash Lite
22M parameters
Ships onto the silicon you already chose
Rivals BADAS 1.0 at 14x fewer params
<3 ms A100 · 5.9 ms Thor
Latency across silicon
Model
Params
AP
A100 (FP16)
Jetson Thor
BADAS 2.0
300M
99.4%
34 ms
41 ms
BADAS 2.0 Flash
86M
99.0%
4.8 ms
12.5 ms
BADAS 2.0 Flash Lite
22M
98.4%
<3 ms
5.9 ms
Flash Lite gives up 1 point of AP at 5.9 ms on Jetson Thor – same architecture and same training as the 300M cloud model.
06 – The Open Challenge
0.994 AP vs 0.940 on the same data.
We trained NVIDIA's COSMOS-Reason2-2B on the exact same 2M clips BADAS 2.0 was trained on, and published the result. BADAS 2.0 lands 99.4% AP at up to 91x fewer parameters.
Metric
BADAS 2.0
COSMOS-BADAS
Average Precision
99.4%
94.0%
Early Warning Recall
91.3%
48.3%
Architecture
V-JEPA2 (Attention)
Autoregressive
Smallest Model
22M params
2B params (cloud only)
Training Data
2M real-world clips
2M real-world clips (same)
Explainability
Native attention maps
None
COSMOS-BADAS = NVIDIA COSMOS-Reason2-2B fine-tuned on the same 2M Empiric Earth training clips used by BADAS 2.0.
The Efficiency Gap: Fine-tuning improved COSMOS by +18.5 pp AP – but at 2B parameters, it still sits 5.4 pp below BADAS 2.0 (94.0% vs 99.4%). BADAS 2.0 Flash Lite (22M) outperforms COSMOS-BADAS (2,000M) by +4.4 pp AP while being 91x smaller.
BADAS 2.0 fine-tunes V-JEPA2, the architecture Yann LeCun proposed for world models. The point isn't the lineage. The point is that latent-space prediction beats pixel reconstruction on collision anticipation, and we publish the benchmarks to prove it.
350K Cameras
Real-world capture 100M+ miles/month
→
Empiric Earth Atlas
GPS-validated pipeline 45 PB structured video
→
V-JEPA2
Self-supervised learning Latent-space prediction
→
BADAS 2.0
94.0% mAP · 4.6% FPR 34ms · up to 91x smaller
BADAS 2.0 fine-tunes a V-JEPA2 ViT-L backbone (300M parameters, 24 transformer layers) end-to-end on 16-frame clips at 256×256 resolution and 8 fps. A future-prediction branch estimates the scene 1 second ahead and concatenates it with the current clip, giving the prediction head access to both present evidence and near-future dynamics. Domain-specific SSL pre-training on 2.25M unlabeled Empiric Earth edge device clips is the critical enabler for the distilled edge variants.
Why V-JEPA2 matters: V-JEPA2 learns by predicting the latent-space representation of future video frames rather than reconstructing pixels. Pixel reconstruction optimizes for visual fidelity. Latent-space prediction optimizes for physical causality. For collision anticipation, you need a model that understands what will happen – not just what is happening.
LATENT-SPACE PREDICTION
178,500 Labeled Videos. Zero Synthetic Data.
Most collision anticipation models are trained on synthetic data or small academic datasets. BADAS 2.0 is trained exclusively on real-world edge device video from Empiric Earth's network – the largest ego-centric driving dataset ever assembled for this task.
BADAS 2.0 is trained on 178,500 labeled videos (~2M windowed clips) – a 4.5x expansion over v1.0. The corpus is assembled through intelligent data mining: BADAS 1.0 runs as an active oracle over millions of unlabeled Empiric Earth drives, surfacing high-risk clips for human review.
The result: 99.4% AP at 4.6% FPR – a 58% reduction in false alarms over v1.0 on the sliding-window benchmark, with gains across all subgroups including the hardest long-tail categories.
Same Architecture. Better Data. Much Better Results.
1.5K
BADAS Open
~40K
BADAS 1.0
~200K
BADAS 2.0
Long Tail of Driving
Excels on rare, edge-case scenarios – animals, fog, snow, motorcyclists, infrastructure failures. 99.4% AP across all 10 long-tail categories where competitors collapse.
Physics, Not Pattern Matching
V-JEPA2 predicts latent-space representations of future frames. This optimizes for physical causality – what will happen – not visual similarity to training data.
Per-Category Dominance
BADAS 2.0 vs COSMOS across all 10 long-tail categories. BADAS 2.0 leads in every single one.
BADAS 2.0
COSMOS-BADAS
COSMOS-Reason2
Models don't emerge from abstractions alone. They come from sustained exposure to reality.
– Yann LeCun, Turing Award Winner, Empiric Earth Board Member
08 – FAQ
Frequently Asked Questions
BADAS 2.0 is Empiric Earth's collision anticipation system, now in its second generation. BADAS 2.0 fine-tunes V-JEPA2 on 178,500 labeled edge device videos (~2M windowed clips) and achieves state-of-the-art accuracy across the public benchmarks we evaluated against.
BADAS 2.0 was evaluated on the Nexar Kaggle competition (1,344 clips, single window), a new 10-group long-tail benchmark (888 clips covering animal, pedestrian, cyclist, fog, rain, snow, intersection, infrastructure, passing/overtaking, and motorcyclist scenarios), and three public external benchmarks: DAD, DoTA, and DADA-2000 using ego-centric re-annotation and sliding-window evaluation.
The paper compares five BADAS variants (2.0, 1.0, 2.0 Flash, 2.0 Flash Lite, Open) against four VLM baselines: Gemini-BADAS (Gemini 2.5 Pro fine-tuned on BADAS data), COSMOS-BADAS (NVIDIA COSMOS-Reason2-2B fine-tuned), vanilla Gemini 2.5 Pro, and Qwen3-VL-2B. Even after fine-tuning on the same data, autoregressive VLMs remain significantly below the BADAS family on the long-tail benchmark.
BADAS 2.0 fine-tunes V-JEPA2 (ViT-L, 300M parameters) end-to-end on edge device video. A future-prediction branch estimates the scene 1 second ahead, giving the classifier access to both present and anticipated dynamics. The distilled variants – BADAS 2.0 Flash at 86M (4x compression) and BADAS 2.0 Flash Lite at 22M (14x compression) – use domain-specific SSL pre-training followed by knowledge distillation to achieve near-parity accuracy at 7–12x faster inference.
BADAS 2.0 is available via a public API that lets you upload video, run predictions, and export results – including prediction overlays and attention heatmaps. Try the web-based playground at the top of this page for direct browser access. Enterprise partners can request full API access for integration into AV programs, fleets, or ADAS platforms.
09 – The Signal
22M parameters. 100M real-world miles a month. Drop it onto your machine.
Upload a clip and see what BADAS 2.0 predicts on your environment. Or book a 30-minute scoping call – we'll walk through how it lands on the silicon you've already shipped.
