QUYVN/BQP

Files

2026-07-13 09:25:40 +07:00

24 KiB

Raw Blame History

CONFIGURATION, WORKFLOWS & IMPLEMENTATION GUIDE

Document: Part 5 of Robot Tuning System Architecture
Coverage: Default configurations, tuning workflows, implementation phases, and testing strategy

DEFAULT CONFIGURATIONS

All default values with detailed justifications.

1. Robot Physical Configuration

public static class DefaultConfigurations
{
    public static RobotPhysicalConfig Physical => new()
    {
        // Wheelbase: Distance between left and right wheels
        // Typical for small indoor robot: 0.3-0.5m
        // Affects: Turning radius, stability
        Wheelbase = 0.35f,  // meters
        
        // Wheel radius: Affects odometry calculations
        // Typical: 0.05-0.1m for small robots
        WheelRadius = 0.075f,  // meters
        
        // Max linear velocity: User-specified
        MaxLinearVelocity = 1.5f,  // m/s
        
        // Max angular velocity: Calculated from max linear velocity
        // ω_max ≈ 2 * v_max / wheelbase
        // Conservative estimate: 2 * 1.5 / 0.35 ≈ 8.57 rad/s
        // Use 6 rad/s for safety margin (~343°/s)
        MaxAngularVelocity = 6.0f,  // rad/s
        
        // Max linear acceleration: For smooth motion
        // 1.0 m/s² means 0 → 1.5 m/s in 1.5 seconds
        // Typical range: 0.5-2.0 m/s²
        MaxLinearAcceleration = 1.0f,  // m/s²
        
        // Max angular acceleration: User-specified
        MaxAngularAcceleration = 1.0f,  // rad/s²
        
        // Robot mass: Estimated for small mobile robot
        Mass = 25.0f  // kg
    };
}

2. Control Timing Configuration

public static ControlTimingConfig Timing => new()
{
    // Control loop frequency: 50Hz (20ms per cycle)
    // Justification:
    // - 10Hz: Too slow, robot will oscillate
    // - 50Hz: Optimal for indoor navigation (balance performance/CPU)
    // - 100Hz: Better but requires more CPU, marginal gains
    // - 200Hz+: Overkill for this application
    ControlLoopFrequency = 50,  // Hz
    
    // Encoder sampling: Match or exceed control frequency
    EncoderSamplingRate = 50,  // Hz
    
    // Motor command rate: Match control frequency
    MotorCommandRate = 50  // Hz
};

3. PID Controller Configuration

public static VelocityPIDConfig PID => new()
{
    // Kp: Proportional gain
    // Higher Kp = faster response but more overshoot
    // Starting point: 0.8 (moderate response)
    // Tuning range: 0.1-5.0
    Kp = 0.8f,
    
    // Ki: Integral gain
    // Eliminates steady-state error
    // Keep small to avoid windup
    // Starting point: 0.1
    // Tuning range: 0.0-2.0
    Ki = 0.1f,
    
    // Kd: Derivative gain
    // Dampens oscillations, smooths response
    // Starting point: 0.05 (gentle damping)
    // Tuning range: 0.0-1.0
    Kd = 0.05f,
    
    // Velocity limits
    MaxVelocity = 1.5f,  // m/s (from physical config)
    MinVelocity = 0.1f,  // m/s (minimum for robot to move)
    
    // Anti-windup: Prevent integral term from growing unbounded
    IntegralWindupLimit = 0.5f,  // m/s
    
    // Output saturation: Ensure output stays within limits
    OutputSaturationEnabled = true
};

PID Tuning Guidelines:

Start with Kp only (Ki=0, Kd=0), increase until oscillation
Reduce Kp to 60% of oscillation value
Add Kd to dampen remaining oscillation
Add Ki last, only if steady-state error exists

4. Velocity Estimator Configuration

public static VelocityEstimatorConfig Estimator => new()
{
    // Alpha filter: Exponential moving average for encoder
    // Lower α (0.1-0.2): More filtering, more lag
    // Higher α (0.3-0.4): Less filtering, more responsive
    // Recommended: 0.3 for balance
    AlphaFilter = 0.3f,
    
    // Blend ratio bounds
    MinBlendRatio = 0.15f,  // Min 15% model, 85% encoder
    MaxBlendRatio = 0.8f,   // Max 80% model, 20% encoder
    DefaultBlendRatio = 0.6f,  // Start with 60% model
    
    // Adaptive blending thresholds
    GoodTrackingThreshold = 0.12f,  // < 12% error
    ModerateTrackingThreshold = 0.3f,  // < 30% error
    
    // CORRECTED blend ratios (based on architecture review):
    // Good tracking → trust encoder more (model prediction matches reality)
    GoodTrackingBlend = 0.3f,  // 30% model, 70% encoder
    
    // Moderate tracking → balanced
    ModerateTrackingBlend = 0.5f,  // 50-50 blend
    
    // Poor tracking → trust model more (encoder may have slip)
    PoorTrackingBlend = 0.7f,  // 70% model, 30% encoder
    
    // Confidence decay: How fast confidence drops
    // 0.98 = slow decay (1% per cycle at 50Hz = ~2s to halve)
    // 0.95 = medium decay
    // 0.90 = fast decay
    ConfidenceDecayRate = 0.98f,
    
    // Minimum confidence floor
    MinConfidence = 0.3f  // Never go below 30%
};

5. Pure Pursuit Configuration

public static PurePursuitConfig PurePursuit => new()
{
    // Lookahead minimum: Smallest lookahead distance
    // Too small: Oscillation, overshoot corners
    // Too large: Cuts corners, poor tracking
    // Recommended: 0.2-0.4m for indoor robot
    LookaheadMin = 0.3f,  // meters
    
    // Kdd: Lookahead velocity scaling factor
    // lookahead = LookaheadMin + Kdd * |velocity|
    // Kdd = 1.0 means 1 second lookahead time
    // Kdd = 0.5 means 0.5 second lookahead
    // Recommended: 0.8-1.5
    Kdd = 1.0f,  // seconds
    
    // Lookahead maximum: Cap lookahead distance
    // Prevents looking too far ahead at high speeds
    // Recommended: 1.5-3.0m
    LookaheadMax = 2.0f  // meters
};

Pure Pursuit Tuning Guidelines:

Increase Kdd for smoother, more predictive tracking
Decrease Kdd for tighter, more reactive tracking
Increase LookaheadMin if robot oscillates
Decrease LookaheadMin if robot cuts corners

6. Path Following Configuration

public static PathFollowingConfig PathFollowing => new()
{
    // Waypoint tolerance: How close to consider waypoint "reached"
    WaypointTolerance = 0.15f,  // 15cm
    
    // Final goal tolerance: Tighter tolerance for final goal
    FinalGoalTolerance = 0.05f,  // 5cm
    
    // Goal heading tolerance: Acceptable heading error at goal
    GoalHeadingTolerance = 5f * MathF.PI / 180f,  // 5 degrees
    
    // Stop distance: When to start preparing to stop
    StopDistance = 0.1f,  // 10cm before goal
    
    // Stop velocity: Threshold to consider robot "stopped"
    StopVelocity = 0.05f  // 5cm/s
};

7. Safety Configuration

public static SafetyConfig Safety => new()
{
    // Max cross-track error before abort
    // 0.5m is reasonable for 10x20m indoor space
    MaxCrossTrackError = 0.5f,  // meters
    
    // Max heading error before abort
    // 45° means robot is severely off course
    MaxHeadingError = 45f * MathF.PI / 180f,  // radians
    
    // Sustained error duration before abort
    // 3 seconds allows recovery from temporary issues
    MaxTrackingErrorDuration = 3000,  // milliseconds
    
    // Obstacle safety distances (for future sensors)
    MinObstacleDistance = 0.3f,  // 30cm emergency stop
    SafetyStopDistance = 0.5f,   // 50cm slow down
    
    // Emergency deceleration limit
    EmergencyStopDeceleration = 2.0f  // m/s²
};

8. Acceptance Criteria Configuration

public static AcceptanceCriteria Criteria => new()
{
    // PRIMARY: Tracking Accuracy (50% weight)
    // Cross-track error RMS: 10cm is good for indoor robot
    MaxCrossTrackErrorRMS = 0.10f,  // meters
    
    // Peak CTE: Allow double RMS as occasional spike
    MaxCrossTrackErrorPeak = 0.20f,  // meters
    
    // Heading error: 10° RMS is acceptable
    MaxHeadingErrorRMS = 10f * MathF.PI / 180f,  // radians
    
    // Goal position error: 5cm final accuracy
    MaxGoalPositionError = 0.05f,  // meters
    
    // Goal heading error: 5° final accuracy
    MaxGoalHeadingError = 5f * MathF.PI / 180f,  // radians
    
    // SECONDARY: Smoothness (30% weight)
    // Max jerk: 5 m/s³ is smooth for human comfort
    MaxJerk = 5.0f,  // m/s³
    
    // Max angular jerk: 10 rad/s³
    MaxAngularJerk = 10.0f,  // rad/s³
    
    // TERTIARY: Efficiency (20% weight)
    // Path length ratio: <15% deviation from optimal
    MaxPathLengthRatio = 1.15f,  // 115% of optimal
    
    // Success rate: 90% of runs should pass
    MinSuccessRate = 0.90f  // 90%
};

9. Scoring Weights

public static ScoringWeights Weights => new()
{
    // How much each category contributes to overall score
    TrackingAccuracy = 0.5f,  // 50%
    Smoothness = 0.3f,        // 30%
    Efficiency = 0.2f         // 20%
};

10. Parameter Bounds

public static class ParameterBounds
{
    // PID bounds
    public static Range KpRange = new(0.1f, 5.0f);
    public static Range KiRange = new(0.0f, 2.0f);
    public static Range KdRange = new(0.0f, 1.0f);
    
    // Estimator bounds
    public static Range AlphaFilterRange = new(0.05f, 0.5f);
    public static Range BlendRatioRange = new(0.1f, 0.9f);
    public static Range ConfidenceDecayRange = new(0.90f, 0.99f);
    
    // Pure Pursuit bounds
    public static Range KddRange = new(0.3f, 2.0f);
    public static Range LookaheadMinRange = new(0.1f, 0.5f);
    public static Range LookaheadMaxRange = new(0.5f, 3.0f);
    
    // Validation rules
    public static List<ValidationRule> Rules => new()
    {
        new ValidationRule
        {
            Name = "LookaheadOrdering",
            Check = (p) => p.PurePursuit.LookaheadMax > p.PurePursuit.LookaheadMin,
            Message = "LookaheadMax must be greater than LookaheadMin"
        },
        new ValidationRule
        {
            Name = "BlendRatioOrdering",
            Check = (p) => p.Estimator.GoodTrackingBlend <= p.Estimator.PoorTrackingBlend,
            Message = "GoodTrackingBlend should be less than PoorTrackingBlend"
        },
        new ValidationRule
        {
            Name = "VelocityLimit",
            Check = (p) => p.PID.MaxVelocity <= p.Physical.MaxLinearVelocity,
            Message = "PID MaxVelocity cannot exceed physical limit"
        }
    };
}

TUNING WORKFLOWS

Detailed step-by-step workflows for different tuning scenarios.

Workflow 1: Quick Start (First-Time User)

Goal: Get robot moving with default settings and validate basic functionality.

Steps:

Load Default Configuration (5 min)
- Open dashboard
- Navigate to Configuration → Robot Settings
- Verify physical parameters (wheelbase, wheel radius)
- Click "Load Default Preset"
Run Baseline Test (2 min)
- Select Test Scenario: "Straight Line 10m"
- Click "Run Test"
- Observe real-time visualization
- Wait for completion
Review Results (3 min)
- Check overall score
- Identify which metrics fail (if any)
- Note: CTE RMS, jerk, smoothness
Decision Point:
- Score > 80: Proceed to Workflow 2 (test other trajectories)
- Score 60-80: Proceed to Workflow 3 (manual tuning)
- Score < 60: Check robot hardware, retry

Expected Outcome: Baseline performance established, ready for tuning.

Workflow 2: Multi-Scenario Validation

Goal: Test current configuration across all trajectory types.

Steps:

Setup Batch Test (2 min)
- Navigate to Tuning → Parameter Comparison
- Select all scenarios:
  - Straight Line 10m
  - Circle 2m Radius
  - Circle 0.5m Radius
- Select current parameter set
- Click "Run Batch"
Monitor Progress (10-15 min)
- Watch each test in sequence
- Note any failures or safety violations
Analyze Comparison (5 min)
- View comparison table
- Identify weakest scenario
- Check metric breakdown per scenario
Decision Point:
- All scenarios pass: Configuration is robust
- One scenario fails: Tune for that specific case
- Multiple scenarios fail: Need general tuning (Workflow 3)

Workflow 3: Manual Iterative Tuning

Goal: Hand-tune parameters to improve specific metrics.

Steps:

Phase 1: Improve Tracking Accuracy (if CTE RMS > 0.10m)

Diagnose Issue:
- View trajectory plot
- Check if robot overshoots or undershoots corners
- Check if error is consistent or oscillating
If Robot Overshoots (cuts corners):
- Decrease Pure Pursuit LookaheadMin: 0.3 → 0.25
- Decrease Kdd: 1.0 → 0.8
- Run test, check improvement
If Robot Undershoots (goes wide):
- Increase Pure Pursuit LookaheadMin: 0.3 → 0.35
- Increase Kdd: 1.0 → 1.2
- Run test, check improvement
If Robot Oscillates:
- Increase PID Kd: 0.05 → 0.15 (more damping)
- Increase Estimator AlphaFilter: 0.3 → 0.4 (more smoothing)
- Run test, check improvement
If Robot is Sluggish:
- Increase PID Kp: 0.8 → 1.2 (faster response)
- Run test, check for overshoot

Phase 2: Improve Smoothness (if Jerk > 5.0 m/s³)

Increase Damping:
- Increase PID Kd: current → +0.1
- Run test
Smooth Velocity Estimates:
- Decrease Estimator AlphaFilter: 0.3 → 0.2
- Run test
Reduce Aggressiveness:
- Decrease PID Kp: current → -0.2
- Run test

Phase 3: Verify and Save

Run Full Validation:
- Test all scenarios with new parameters
- Ensure no regressions
Save Configuration:
- Name: "Tuned_[Date]_v1"
- Add description of changes
- Click "Save"

Iteration: Repeat phases as needed, aiming for <5 iterations.

Workflow 4: Automated Optimization (Advanced)

Goal: Use algorithm to find optimal parameters automatically.

Steps:

Configure Optimization (5 min)
- Navigate to Tuning → Auto Tuning
- Select algorithm: Bayesian Optimization
- Select parameters to tune:
  - ☑ PID: Kp, Ki, Kd
  - ☑ Pure Pursuit: Kdd
  - ☐ Estimator: (keep fixed for first run)
- Set constraints:
  - Max iterations: 30
  - Early stopping: 1% improvement threshold
Define Objective (2 min)
- Primary metric: Cross-Track Error RMS
- Secondary metric: Max Jerk (weight: 0.3)
- Test scenario: Circle 2m Radius
Start Optimization (30-60 min)
- Click "Start Optimization"
- Monitor progress dashboard
- View live updates of best parameters found
Review Results (10 min)
- Check final parameters
- Compare to baseline
- Review improvement %
Validate on Other Scenarios (15 min)
- Run batch test with optimized parameters
- Ensure no regressions on other trajectories
Save or Iterate:
- If satisfied: Save as "Optimized_v1"
- If not: Adjust weights, re-run optimization

Workflow 5: A/B Testing Configurations

Goal: Compare two parameter sets side-by-side.

Steps:

Select Configurations (2 min)
- Config A: "Default"
- Config B: "Tuned_v1"
Choose Test Scenario (1 min)
- Straight Line 10m
Run Comparison (5 min)
- Click "Run Comparison"
- System runs both tests sequentially
Analyze Results (5 min)
- View side-by-side metrics table
- Check trajectory overlay plot
- Identify winner
Statistical Significance (optional)
- Run 10 trials each
- Compare mean ± std deviation
- Determine if difference is significant

IMPLEMENTATION PHASES

Phased approach to building the system, prioritized by value and complexity.

Phase 0: MVP (Weeks 1-3)

Goal: Core functionality for manual tuning.

Features:

✅ Single test execution
✅ Manual parameter adjustment UI
✅ Real-time visualization (2D trajectory)
✅ Basic metrics calculation
✅ Save/load configurations
✅ Data logging

Deliverables:

Working Blazor dashboard
Integrated controllers (PID, Estimator, Pure Pursuit)
Basic test executor
SQLite database with core tables
Real-time SignalR updates

Tech Stack:

Blazor Server
Entity Framework Core + SQLite
SignalR
Plotly.NET for charts

Testing:

Unit tests for controllers
Integration test for one full test run
Manual UI testing

Success Criteria:

User can run a test and see results
Parameters can be adjusted and re-run
Metrics are calculated correctly

Phase 1: Enhanced Tuning (Weeks 4-6)

Goal: Multi-scenario testing and comparison.

Features:

✅ Batch testing
✅ Configuration comparison
✅ Test history viewer
✅ Safety monitoring with abort
✅ Parameter validation
✅ CSV export

Deliverables:

Batch test executor
Comparison UI components
Enhanced database queries
Safety monitor implementation
Report generation (HTML/CSV)

Testing:

Batch test with 3 scenarios
Comparison test with 3 configs
Safety violation test

Success Criteria:

Batch tests complete without manual intervention
Comparison clearly shows best configuration
Safety system aborts on violations

Phase 2: Advanced Analytics (Weeks 7-9) - OPTIONAL

Goal: Deep insights and semi-automated tuning.

Features:

⭕ Statistical analysis
⭕ Trend analysis over time
⭕ Parameter sensitivity analysis
⭕ Tuning suggestions
⭕ PDF report generation

Deliverables:

Statistics calculator
Trend visualization
Suggestion engine (rule-based)
PDF generator

Testing:

Historical data analysis (50+ tests)
Suggestion accuracy validation

Success Criteria:

Trends clearly visible
Suggestions improve results

Phase 3: Automated Optimization (Weeks 10-13) - FUTURE

Goal: Hands-off parameter optimization.

Features:

❌ Grid search
❌ Random search
❌ Bayesian optimization
❌ Genetic algorithm
❌ Multi-objective optimization

Deliverables:

Optimization framework
Multiple algorithm implementations
Hyperparameter tuning for optimizers
Parallel evaluation (if multiple robots)

Complexity: Very High (requires ML libraries)

Testing:

Benchmark against manual tuning
Convergence tests
Robustness tests

Success Criteria:

Automated optimization finds better params than manual in <1 hour
Reproducible results

TESTING STRATEGY

Comprehensive testing at all levels.

1. Unit Tests

Coverage Target: >80% for domain logic

Key Tests:

// PID Controller Tests
[Fact]
public void PIDController_ProportionalOnly_CorrectOutput()
{
    var config = new VelocityPIDConfig { Kp = 1.0f, Ki = 0, Kd = 0 };
    var pid = new PIDController(config);
    
    var output = pid.Calculate(error: 1.0f, dt: 0.02f);
    
    Assert.Equal(1.0f, output, precision: 2);
}

[Fact]
public void PIDController_IntegralWindup_Clamped()
{
    var config = new VelocityPIDConfig 
    { 
        Kp = 0, 
        Ki = 1.0f, 
        Kd = 0,
        IntegralWindupLimit = 0.5f
    };
    var pid = new PIDController(config);
    
    // Accumulate large error
    for (int i = 0; i < 100; i++)
        pid.Calculate(error: 1.0f, dt: 0.02f);
    
    var output = pid.Calculate(error: 1.0f, dt: 0.02f);
    
    Assert.True(output <= 0.5f);
}

// Velocity Estimator Tests
[Fact]
public void VelocityEstimator_GoodTracking_TrustsEncoder()
{
    var config = DefaultConfigurations.Estimator;
    var estimator = new VelocityEstimator(config, new VelocitySignalProcessingConfig());
    
    var vHybrid = estimator.EstimateVelocity(
        vCmd: 1.0f, 
        vActual: 0.95f,  // Close to command (good tracking)
        dt: 0.02f
    );
    
    // Should blend more toward encoder (0.95) than model
    Assert.True(vHybrid > 0.93f && vHybrid < 0.97f);
}

// Pure Pursuit Tests
[Fact]
public void PurePursuit_LookaheadScalesWithVelocity()
{
    var config = new PurePursuitConfig 
    { 
        LookaheadMin = 0.3f,
        Kdd = 1.0f,
        LookaheadMax = 2.0f
    };
    var pp = new PurePursuitController(config);
    
    var path = CreateStraightLinePath(10);
    var pose = new Pose2D(0, 0, 0);
    
    pp.Calculate(pose, velocity: 0.5f, confidence: 1.0f, path);
    var lookahead1 = pp.GetLookaheadDistance();
    
    pp.Calculate(pose, velocity: 1.0f, confidence: 1.0f, path);
    var lookahead2 = pp.GetLookaheadDistance();
    
    Assert.True(lookahead2 > lookahead1);
}

2. Integration Tests

[Fact]
public async Task FullTestRun_StraightLine_Completes()
{
    // Arrange
    var scenario = CreateStraightLineScenario(10);
    var parameters = DefaultConfigurations.GetDefaultPreset();
    var orchestrator = CreateOrchestrator();
    
    // Act
    var result = await orchestrator.RunSingleTest(scenario, parameters);
    
    // Assert
    Assert.Equal(TestStatus.Completed, result.Status);
    Assert.NotNull(result.Metrics);
    Assert.True(result.Metrics.CrossTrackErrorRMS < 0.20f);
}

[Fact]
public async Task SafetyMonitor_ExcessiveCTE_AbortsTest()
{
    // Arrange
    var scenario = CreateCircleScenario(0.5f); // Tight circle
    var parameters = CreateBadParameters(); // Intentionally bad
    var orchestrator = CreateOrchestrator();
    
    // Act
    var result = await orchestrator.RunSingleTest(scenario, parameters);
    
    // Assert
    Assert.Equal(TestStatus.Aborted, result.Status);
    Assert.True(result.SafetyViolations.Any(v => v.Type == ViolationType.CrossTrackError));
}

3. Performance Tests

[Fact]
public void ControlLoop_MaintainsFrequency()
{
    var executor = CreateTestExecutor();
    var scenario = CreateStraightLineScenario(5);
    var parameters = DefaultConfigurations.GetDefaultPreset();
    
    var timestamps = new List<long>();
    
    executor.ExecuteAsync(
        onStateUpdate: state => timestamps.Add(state.TimestampMs)
    ).Wait();
    
    // Calculate actual frequency
    var intervals = timestamps.Zip(timestamps.Skip(1), (a, b) => b - a);
    var avgInterval = intervals.Average();
    var actualFrequency = 1000.0 / avgInterval;
    
    Assert.InRange(actualFrequency, 48, 52); // 50Hz ± 2Hz
}

4. End-to-End Tests

Manual Test Plan:

Happy Path Test
- Load default config
- Run straight line test
- Verify metrics displayed
- Save configuration
- Reload and verify
Error Handling Test
- Set invalid parameter (Kp = 100)
- Attempt to run test
- Verify error message shown
- Verify test doesn't start
Real-time Update Test
- Start test
- Verify UI updates at ~10Hz
- Pause test
- Verify pause works
- Resume and complete
Comparison Test
- Create 2 configs
- Run comparison
- Verify side-by-side display
- Export results

DEPLOYMENT CHECKLIST

Pre-deployment validation:

All unit tests pass
Integration tests pass
Performance tests pass
Manual E2E tests completed
Database migrations created
Default data seeded
Configuration files reviewed
Robot hardware tested
Emergency stop tested
Documentation complete
User manual created

This completes the Configuration, Workflows & Implementation Guide.

24 KiB Raw Blame History Unescape Escape

CONFIGURATION, WORKFLOWS & IMPLEMENTATION GUIDE

DEFAULT CONFIGURATIONS

1. Robot Physical Configuration

2. Control Timing Configuration

3. PID Controller Configuration

4. Velocity Estimator Configuration

5. Pure Pursuit Configuration

6. Path Following Configuration

7. Safety Configuration

8. Acceptance Criteria Configuration

9. Scoring Weights

10. Parameter Bounds

TUNING WORKFLOWS

Workflow 1: Quick Start (First-Time User)

Workflow 2: Multi-Scenario Validation

Workflow 3: Manual Iterative Tuning

Workflow 4: Automated Optimization (Advanced)

Workflow 5: A/B Testing Configurations

IMPLEMENTATION PHASES

Phase 0: MVP (Weeks 1-3)

Phase 1: Enhanced Tuning (Weeks 4-6)

Phase 2: Advanced Analytics (Weeks 7-9) - OPTIONAL

Phase 3: Automated Optimization (Weeks 10-13) - FUTURE

TESTING STRATEGY

1. Unit Tests

2. Integration Tests

3. Performance Tests

4. End-to-End Tests

DEPLOYMENT CHECKLIST

24 KiB

Raw Blame History