AGENTS

Autonomous Driving Log Analysis Skill

Dedicated skill for ADAS algorithm scenarios like parking-spot optimization: log parsing, AI analysis, code edits and validation.

Core Pain Points

1. Raw logs are hard to analyze directly: Real-scene pls.txt log files contain a large volume of raw records; manual parsing, filtering, and anomaly localization is time-consuming.
2. Complex relationship between NG root cause and algorithms: NG defect scenarios in the validation dataset need to be compared against the general test dataset; judging which algorithm was called and the root cause depends on domain expertise.
3. Long modify-validate chain: Multiple round-trips exist from problem analysis to code modification, human review, and validation-set verification.

Solution & Technical Implementation

1. Standardized input: Accepts validation datasets, NG defect-scenario data, general test datasets, and real-scene raw pls.txt logs.
2. Dedicated skill orchestration: Uses skill-creator to build new skills supporting algorithm log parsing, code modification, and result visualization.
3. Analyze-modify-validate closed loop: Sequentially performs log parsing, AI analysis of NG root cause, algorithm invocation, human review of AI suggestions, AI code modification, and validation-set verification.

Highlights

• Customizable domain skills: Builds dedicated capabilities for specific algorithm scenarios such as parking-slot optimization.
• Direct raw-log connection: Processes pls.txt directly, reducing manual preprocessing.
• Preserves human quality gates: AI suggestions and code modifications undergo human review before validation.

Business Value

• Reduces time spent organizing raw logs and analyzing NG root causes, improving algorithm-debugging efficiency.
• The skill-creation approach can be ported to other AD algorithms, data formats, and defect types.
• Turns log parsing, algorithm invocation, and validation steps into a fixed workflow, reducing dependence on a small number of experts.

---