How The AI Institute is Revolutionizing Robotics ML Training
summary
Boston Dynamics AI Institute ships an end-to-end Python platform called RobotTensors platform that glues Ray Data, Ray Train, Ray Tune and RayClusters to GKE storage and Weights & Biases for model registry and metrics. It trains a diffusion policy on a 12-task mimicgen dataset, evaluates asynchronously every 10 epochs by autoscaling a separate simulation pool, and runs thousands of simulations per training run to return success rates. Multi-tenancy uses per-user dev Ray clusters and a transient-prod queue; data lives as HDF5/parquet on GCS and the system caught a preprocessing bug that broke episode starts.
AI-written summary · about
chapters (10)
- 0:00Introduction to AI Institute
- 1:28Robotics ML training challenges
- 6:46Robot Tensors platform overview
- 9:16Multi-tenant cluster design choices
- 11:04Simulation evaluation in training loop
- 13:11Demo setup and custom cluster creation
- 16:04Training diffusion policy model
- 21:15Debugging and data preprocessing
- 22:39Production job submission and scaling
- 24:25Q&A session
related talks
25mAI In Production Key Challenges and Use Cases
31mHosting AI models with Ray and GKE
35mOperating a Production-Scale Physical AI Data Platform with Ray
28mScaling AI & Machine Learning Workloads With Ray on AWS, Kubernetes, & BERT
21mHow Ray and Anyscale Make it Easy to do Massive-scale ML on Aerial Imagery
28mScaling Multimodal Data and Reinforcement Learning with Ray
22mRay, a Unified Distributed Framework for the Modern AI Stack
38mAI in Kubernetes: How to Get Started?