TODO.md

Plan

Foundations

• Smooth the workflow
  • Find a good linear algebra library
  • Write operator overloads, extensions
  • Test EJML and extensions
  • Translate all src to Kotlin
  • Remove Mat2 Vec2 Vec3
  • Find command line compile and run commands
• Reproduce HW2 EKF SLAM in java using processing.
• Replace DMatrix2 with FMatrix2

Simulator

• significance of time in odometry and laser scan struct
• scaling
• 3d rendering

Obstacle and landmark extraction

• Extract obstacles
  • RANSAC/LS
    • Return end points not defining points
    • Line fitting
    • Fix partition code
    • use IEP for better partitioning
  • IEP
• Extract landmarks
  • No need for checking intersections separately due to IEP
  • Loose end detection in first stage
  • Intersection detection in second stage
• Advanced landmark detection
  • Intersection landmarks
    • use perpendicular projections points
    • check distances (IEP shortcoming)
  • Loose endpoints
    • Almost parallel to a long wall case. Just increase DISCONTINUITY THREASHOLD
    • Simple block bug

Estimation

• Estimate Sigma_N
  • Give a control for some time and not errors, fit a gaussian with that
• Estimate Sigma_M
  • Sigma_M needs the RANSAC/LS/Corner detector

SLAM

• SLAM
  • Propogation
  • Augement
  • Update
  • Do all updates first and then augment all new landmarks
  • Periodically clean landmarks which have very less number of hits
  • Use estimated sigmas

Planning

• HitGrid
  • Keep track of obstacles detected
  • Is long needed for count? No.
  • PRM/A* on Hitgrid
  • Control
  • Path smoothing

Smooth differential agent

• Handle smooth differential drive agent. Maybe just use small speeds?

Observations

• SLAM is not as transparent as first half due to matrix gymnastics
• Landmarks and obstacles are not the same
  • Here corners are landmarks and stored in slam state
  • Lines are obstacles, endpoints stored only for planning
• Use time elapsed for robot to calculate dt for propagation, augment and update
• Ignore measurements while moving, do it only while stationary? Moving slowly is fine
• SLAM and planning can be developed independently
• Policy
  • Initially no noise assumption, No more noise while staying stationary
  • 0 control -> doesn't add more drift
  • While true
    • Come to stop and measure
    • Find landmarks with good uncertainity (augment/update)
    • Find obstacles with good uncertainity
    • Plan a path
    • Small valued controls -> less noise, move very slowly (propogate)

Report and demos

• RK4 vs Eular stability vs dt graph
• Estimating Sigma_m does it depend on distance?
• Estimating Sigma_n does it depend on controls?
• Estimating Sigma_m while stationary only if we are taking measurements while stationary
• Check the command line compilation
• Tune params

Optimizations

• Use proper += -= implementations
• Merge A * B^T operation as one
• Don't create new memory for get block??

Future work

• Circular obstacles