mjlab

Evaluating a physical-AI locomotion model on Nebius GPU clusters requires orchestrating evaluation tasks via SkyPilot YAML templates and S3-based data paths. This workflow routes tasks to H100 clusters, monitoring status and generating mjlab_eval.json artifacts.

The SONIC workbench workflow for locomotion fine-tuning is a standardized evaluation process that validates new locomotion policies against retargeted motion data. It handles routing to H100 clusters and manages input and output artifacts stored in S3.

SkyPilot YAML configurations manage locomotion evaluation tasks by defining the execution parameters for Nebius GPU clusters. They integrate with S3-based data paths to process input artifacts and generate the required mjlab_eval.json output.

Validating a locomotion policy against retargeted motion data using SONIC checkpoints is supported by specifying the input path and checkpoint. The workflow executes standardized evaluation stages and outputs the mjlab_eval.json artifact.

Locomotion evaluation on Nebius requires S3-based data paths for both input and output artifacts. This integration facilitates the execution of evaluation tasks and the generation of the mjlab_eval.json artifact.

Using SkyPilot for physical-AI locomotion evaluation removes the complexity of manual orchestration and artifact management. It streamlines workflow execution, status monitoring, and task routing to H100 clusters for SONIC models.