Simulating the Physical World at Scale: Humanoid Locomotion with Ray
summary
Lambda uses Ray and NVIDIA Isaac Sim to scale Unitree G1 gait fine‑tuning across a two‑node Lambda GPU cluster (16 GPUs) running 512 environments per GPU for a total of 8,192 parallel sims and about 139k environment steps per run. Training runs headless, map env generation and gradient compute to each GPU, and reduce fall rate from ~4.75% to ~0.78% while increasing distance per episode, showing the fine‑tuned policy walks farther without cheating on humanoid constraints.
AI-written summary · about
chapters (7)
- 0:00Introduction and background
- 1:31Robot gait improvement setup
- 4:41Training environment details
- 6:47Human-like movement constraints
- 8:12Ray-based distributed training
- 9:36Fine-tuning results and validation
- 11:25Final demonstration and summary
related talks
31mFaster Model Serving with Ray and Anyscale
28mEfficient Deep Learning Training with Ludwig AutoML, Ray, and Nodeless Kubernetes
29mScaling ML Pipelines for Autonomous Driving with Ray
28mScaling Multimodal Data and Reinforcement Learning with Ray
32mAnyscale's Ray Data: Revolutionizing Batch Inference
21mHow Ray and Anyscale Make it Easy to do Massive-scale ML on Aerial Imagery
38mNetflix's Ray Platform: From Deep Learning to GenAI
26mA Growing Ecosystem of Scalable ML Libraries on Ray