Keynote: Fueling Innovation Engines by Creating, Collecting, and Improving Stepping...
summary
Claims that open-ended innovation needs automatic goal switching and diverse stepping stones, backed by concrete tools and results: quality-diversity search using MAP-Elites and population methods, an innovation engines system plus Bayesian optimization for repairs, and the POET algorithm that co-evolves environments and agents. Experiments show a damaged robot relearns walking in about 30–90 seconds from a simulated archive, and Montezuma’s Revenge scores jump to as high as 18 million, outperforming prior approaches. Code is released open source.
AI-written summary · about
chapters (19)
- 0:00Intro
- 0:21A Paradox
- 1:05Key for Science, Technological & Evolutionary Innovation: Generating Problems, Goal Switching
- 5:07Quality Diversity Algorithms
- 6:09Qualitatively Different
- 7:04Goal Switching is Essential for Success
- 7:58Innovation Engines
- 9:14Goal Switching Enables Good Ideas to Spread
- 10:37nature
- 11:12Go-Explore: Solves Montezuma's Revenge
- 12:01What's missing?
- 12:59Paired Open-Ended Trailblazer (POET)
- 13:29Task: Obstacle Courses
- 14:29Another Challenging Environment
- 15:00Direct Path Curriculum Fails
- 16:36Goal Switching Needed to solve the Hardest Challenges
- 17:23Al-Generating Algorithms
- 19:16The open-source community acts like a QD algorithm
- 19:58More Open Source Projects at Uber
related talks
11mKeynote: Open Source Stack Empowers the AI Era
16mKeynote: Marin: An Open Lab for Frontier AI
45mKeynote: Writing Code? Pfft... Evolve it Instead!
49mBeyond GenAI
14mEdge of the Present: Pursuing AI’s Frontier in the Enterprise
46mKeynote: Make your car self-driving using open-source software
34mKeynote: Building for AI at Scale: Infra, Integrity, & Innovation
21mIdeas that Matter