Scale Robot Policy Evaluation with Ray and Isaac Lab
Deploy the GR00T policy as a Ray Serve deployment, then launch Isaac Lab simulators as Ray tasks to achieve closed‑loop, GPU‑isolated evaluation across a cluster.
Deploy the GR00T policy as a Ray Serve deployment, then launch Isaac Lab simulators as Ray tasks to scale closed‑loop evaluation.
Summary
NVIDIA’s GR00T‑N1.7‑3B vision‑language‑action model is served via Ray Serve as an HTTP microservice, exposing a /predict endpoint that returns a 40‑step action chunk in roughly 1.6 s. The policy runs on a dedicated GPU per Ray actor, while each Isaac Lab simulator is launched as an isolated Ray task on its own GPU, preventing memory contention between physics rendering and model inference.
Closed‑loop evaluation is achieved by having each simulator query the policy once per action chunk, allowing many rollouts to share a small fleet of policy replicas. Ray Serve’s autoscaling and load‑balancing mean a single deployment can be scaled from one to dozens of replicas with a single line of code. Anyscale packages the heterogeneous stack into a reproducible cluster image, so scaling from a single rollout to 100+ parallel rollouts requires no infrastructure rebuild.
The approach solves three key challenges: keeping evaluation online, separating GPU‑bound workloads, and sharing policy replicas across rollouts. It demonstrates how Ray and Anyscale can turn a heavy GPU simulation and policy inference pipeline into a scalable, managed cluster service.
Key changes
- GR00T‑N1.7‑3B model served via Ray Serve HTTP endpoint
- Simulators run as isolated Ray tasks each on its own GPU
- Policy replicas autoscaled through Ray Serve
- Inference round‑trip ~1.6 s for 40‑step action chunk
- Simulators query policy once per action chunk
- Cluster image packaged by Anyscale for reproducibility
- Policy and simulator communicate over HTTP
- Health‑check timeout extended to 300 s for 3B model load