train-policy

Robot policy training is coordinated end-to-end across frameworks like LeRobot, Isaac Lab, and SONIC using declarative YAML specifications. This unified workflow approach manages planning, execution, and evaluation while integrating with Nebius infrastructure for GPU allocation and S3-based artifact management.

Large-scale robot policy training on Nebius H100 clusters is executed by specifying a target GPU and a workflow YAML file. This leverages Nebius infrastructure for scalable GPU allocation and S3 storage to manage artifacts, ensuring reproducible training workflows.

Yes, the policy training workflow supports diverse environments including LeRobot, Isaac Lab, and SONIC. It coordinates training, evaluation, and export tasks across these frameworks by leveraging NPA workbench tools and declarative YAML specifications.

Managing reproducible machine learning artifacts is best achieved by integrating training workflows with S3-based artifact management. This approach ensures reproducible and scalable training runs by storing outputs alongside declarative YAML configurations and Nebius GPU infrastructure allocations.

Yes, Nebius infrastructure integration is required for GPU allocation and S3-based artifact management to ensure reproducible and scalable training workflows. The skill coordinates the end-to-end lifecycle across multiple frameworks by relying on this infrastructure.

Evaluating and exporting robot policies is handled within the unified workflow management interface. After training execution, the skill coordinates evaluation and export tasks for various robot policy families using declarative YAML specifications.