Deeper Dive on Making Boats Fly with AI on Ray
summary
Industrial flight-like yacht control is built: a commercial simulator is virtualized into a Gym environment, trained with deep Reinforcement Learning using a custom SAC implementation on a Ray cluster with preemptible cloud instances, and hyperparameters managed by Tune. The system matches or exceeds Olympic sailors in-simulator, cuts design iteration times from days to hours, and focuses on reward shaping, replay-buffer management and checkpointing to reduce variability so agents generalize across nearby design variants.
AI-written summary · about
chapters (12)
- 0:00Intro
- 0:45The Americas Cup
- 5:57The Problem
- 10:02The Race Stack
- 14:55What happens in practice
- 17:11Building the right reward function
- 18:34Stability and robustness
- 21:28Reinforcement learning
- 23:06Demonstration
- 24:13Technology
- 24:40Learnings
- 26:07Conclusion
related talks
20mProject Bonsai: How Microsoft is Using Ray and RLlib Inside a Machine Teaching Service
30mTransforming Grab’s AI Platform with Ray
24mAI in maritime - Autonomous ships
48mMachine Takeover - Improving Your Game With Machine Learning & AI
30mAdmiral-Driven ML Framework for Marine Operations
31mLeveraging Agent Skills for Ray Workloads
28mScaling Multimodal Data and Reinforcement Learning with Ray
36mSkyRL: Democratizing RL Training for Researchers