alpamayo1-x-rl

What problem does it solve?

This Skill automates the process of running end-to-end open-loop RL post-training for Alpamayo VLM action models on the Physical AI AV (PAI) dataset using Cosmos-RL + GRPO.

Core Features & Use Cases

• Convert release checkpoint to training-ready format: Converts Alpamayo model checkpoints to a training-ready format.
• Curate PAI mini subset: Extracts a small subset of the PAI dataset for local training.
• Launch RL training: Runs RL training on the dataset using Cosmos-RL and GRPO.
• Export trained policy checkpoint: Converts the trained policy checkpoint to a HuggingFace checkpoint directory for inference.

Quick Start

Use the alpamayo1-x-rl skill to train an Alpamayo 1.5 model using the PAI dataset and export the resulting policy checkpoint to a HuggingFace directory.