1 of 13

​

4 min presentation in DD2438

Group 1:8

Oliver Sveijer and Adam Ekelöf

2 of 13

What we did

  • Kinodynamic KD-RRT*
    • Accounts for complex dynamics
  • Buffered PD-controller
    • Take a weighted average of n future velocities
    • Better for handling braking and sharp turns

3 of 13

What we missed

  • No post processing
  • Only follows the RRT* path
    • Could be optimal but not expected
    • Computation constraints

4 of 13

The plan (old slide, hidden)

  • Kinodynamic RRT*
  • Post process path into optimal driving line for car

Ref. [Penn] L22 Raceline Optimization

Ref. Towards time-optimal race car driving...

5 of 13

Current progress (old slide, hidden)

  • Almost working kinodynamic RRT
    • For both drone and car

6 of 13

RRT sampling

BFS from goal to get distance for each cell

Uniform distribution over all cells within interval of distance values

Move interval towards goal

​

7 of 13

Problems with sampling

Low adaptability, would need higher concentration in tight spaces

High dimensionality (need more iterations or smarter discard)

8 of 13

RRT*

This becomes even more of a problem when we want to find better paths using *

Transitions between higher dimensions require good models.

9 of 13

Improvement ideas

  • Multiple faster passes over sampler, persuasion
  • Simpler path finding w/ post processing

10 of 13

Next steps (old slide, hidden)

  • Make RRT*
  • Post-process

11 of 13

Optimizations (step 1) (old slide, hidden)

  • Greedy

12 of 13

Optimizations (step 2) (old slide, hidden)

13 of 13

Progress Status Week 6

  • Planned Time spent: 70%
  • Actual Time spent: 80%
  • Actual Progress: 76%
  • Risk of not completing assignment: 11%
    • Not failing the course… Yet.

​