Driving Analysis with Chain-of-Thought

Autonomous driving safety analysis using dashcam video with step-by-step reasoning.

---

Terminal Recording

📺 Can't see the animation? Download MP4

View full output

$ python examples/03_driving_analysis.py
=== 03: Driving Analysis (CoT) ===
Loading nvidia/Cosmos-Reason2-2B (vision, reasoning=True)... ✅ loaded
Processing video: sample.mp4 @ 4 FPS... 40 frames

Agent:
<think>
Let me analyze this dashcam footage step by step.

First, key elements:
- Suburban residential street
- Multiple parked vehicles on both sides
- A pedestrian on the right sidewalk
- Trees creating partial shade
- Road appears dry, good visibility

Potential hazards:
1. Pedestrian could step into the road
2. Car could open a door from parked vehicles
3. Limited visibility at upcoming intersection
4. Children might be playing behind parked cars
</think>

⚠️  Safety Hazards:
• Pedestrian on right sidewalk — risk of jaywalking
• Parked cars on both sides — door opening risk
• Upcoming intersection — limited cross-traffic visibility

✅ Recommended Actions:
• Maintain ≤25 mph through residential zone
• Stay center-lane, avoid door zone
• Watch for pedestrian course changes
• Slow to 15 mph approaching intersection

Time: 16.4s
=== PASS ===

Play locally: asciinema play docs/assets/casts/03_driving_analysis.cast

---

Code

examples/03_driving_analysis.py

from strands import Agent
from strands_cosmos import CosmosVisionModel

model = CosmosVisionModel(
    model_id="nvidia/Cosmos-Reason2-2B",
    reasoning=True,                           # Enable <think> CoT
    fps=4,
    params={"max_tokens": 2048, "temperature": 0.6},
)
agent = Agent(model=model)

result = agent(
    "<video>sample.mp4</video> "
    "Identify safety hazards and recommend actions."
)

Chain-of-Thought Flow

graph LR
    V["🎬 Dashcam Video"] --> F["Extract Frames<br/>@ 4 FPS"]
    F --> M["Cosmos-Reason2"]
    M --> T["&lt;think&gt;<br/>Step-by-step<br/>analysis<br/>&lt;/think&gt;"]
    T --> H["⚠️ Hazards"]
    T --> A["✅ Actions"]

    style T fill:#fff3cd,stroke:#ffc107,color:#333
    style H fill:#9b2226,color:#fff
    style A fill:#2d6a4f,color:#fff

Why Chain-of-Thought Matters for Driving

Without CoT, the model might miss subtle hazards. With reasoning=True:

Aspect	Without CoT	With CoT
Output	Direct answer	Reasoning → Answer
Hazard detection	Obvious only	Subtle + obvious
Confidence	"Be careful"	Specific actions with rationale
Token count	~200	~600
Inference time	~8s	~16s

Built-in driving prompt

Cosmos includes an optimized driving analysis prompt:

from strands_cosmos.cosmos_vision_model import TASK_PROMPTS
print(TASK_PROMPTS["driving"])
# "The video depicts the observation from the vehicle's camera.
#  You need to think step by step and identify the objects in
#  the scene that are critical for safe navigation."

Use Cases

• Fleet safety monitoring — Automated dashcam review
• ADAS development — Scene understanding for driver assistance
• Insurance analysis — Incident reconstruction from video
• Driving instruction — Identify learning moments in student footage
• Autonomous vehicle testing — Perception validation

---

→ Next: Embodied Reasoning | All Examples