Environmentally Conscious AI: Improving Spatial Analysis and Reasoning
•
2 gefällt mir•955 views
This presentation was presented during GDC 2014, AI Summit. In this talk was presented a unified architecture for spatial reasoning using a database-like interface. The objective is to be able to simply query and manipulate spatial information in autonomous characters behaviors.
Empfohlen
Empfohlen
Weitere ähnliche Inhalte
Ähnlich wie Environmentally Conscious AI: Improving Spatial Analysis and Reasoning
Ähnlich wie Environmentally Conscious AI: Improving Spatial Analysis and Reasoning (20)
Kürzlich hochgeladen
Kürzlich hochgeladen (20)
Environmentally Conscious AI: Improving Spatial Analysis and Reasoning
- 1. Environmentally Conscious AI Improving Spatial Analysis and Reasoning Clodéric Mars MASA LIFE Lead cloderic.mars@masagroup.net @clodericmars
- 2. What is Spatial Reasoning? •Spatial objects • POIs • Areas/zones • Influence map • ... • Relationships • Distance • Visibility • Accessibility • ... !2
- 3. “What is the closest bone that is not too old, out of the cat likely position and that I can chew while observing the street?” !3
- 4. How? •List of points of interests •Query several sources • “closest” ➡ pathfinding • “not too old” ➡ metadata • “likely position” ➡ influence map • “observing” ➡ visibility mesh, pvs •Add some custom logic ! ➥ A little complex for “just” a dog! !4
- 5. Wouldn’t it be cool if... •Behavior designers can focus on using • “One stop” gateway • Expressive query •Developers can extend and optimize ! ➥ Handle the “technical” complexity for the benefit of both !5
- 6. Inspirations •Game industry • Environment Tactical Querying (ETQ) [Zielinsky 2013] • Tactical Position Selection (TPS) [Jack 2013] •GIS • PostGIS (postgis.net) !6
- 7. !7
- 8. Let’s study a practical example! !8
- 9. Tactical behaviors in VBS2 •Tactical groups... •...navigating in a “hostile” environment •Behaviors that don’t depend on the environment !9
- 10. !10
- 11. “select cover” query SELECT p FROM interest_points WHERE p.metadata.type == 2 AND (p.metadata.assignee == NULL OR p.metadata.assignee == $.id) AND p.geometry.x > $.position.x - 10 AND p.geometry.x < $.position.x + 10 AND p.geometry.y > $.position.y - 10 AND p.geometry.y < $.position.y + 10 ORDERED BY distance(p.geometry, $.position) LIMIT 10 collection type (2 is cover points) free or already assigned to me no farther than 10 m order by distance to me take the first 10 !12
- 12. Reach a zone and take cover •“select cover” • SpatialDB query • Distance • Protected direction • Availability •“assign cover” • SpatialDB update • Transactional goto cover goto zone close to zone? select cover assign cover !13
- 13. Technical stuffs •SQLite (www.sqlite.org) • Easy to integrate • Public Domain • Mature • Fast • Geospatial index (r-tree) • Functions • In-Memory •peg/leg (piumarta.com/software/peg/) !14
- 14. !15
- 15. How? … with this implementation •List of points of interests (bones) ! •Query several sources (navmesh, visibility, …) ! •Add some custom logic !16 ‣SQLite Tables • metadata • spatial index ‣SQLite Functions • Signature • Callback to subsystem • Metadata access/index ‣SQLite ‘arithmetic’ and custom parser
- 16. What worked •Fast implementation of first version •Data driven behaviors •Familiar query language •OK performances !17 "select cover" duration 0,08 ms 0,15 ms 0,23 ms 0,30 ms 0 1250 2500 3750 5000 Core i7 @ 2Ghz // 8Go of RAM Total size 3 000 kb 6 000 kb 9 000 kb 12 000 kb 0 1250 2500 3750 5000
- 17. What needs to be improved •Easy-to-use assignment •Tools • inspection • sandbox •Performances • grouping/ordering evaluators • time slicing • multithreading !18
- 18. Takeaways !19 1.Decouple (designer/developer) 2.Use existing tech (eg. SQLite) 3.Create a familiar language Meet us at booth #1743 @masalife_ai