ethan-li-k/astra_bench/skill-mint

skill_mint — "earn your place in the catalogue"

Overview

This is the skill that closes the flywheel:

1. You collect rollouts under some ad-hoc label (say handover_block_wip).
2. Once enough successes have piled up, skill_mint promotes that label

into a real task skill directory with SKILL.md + bundle.yaml.

1. After minting, roborsi task list / roborsi plan see the new

task and can use it in long-horizon plans.

The minting step is on purpose explicit — we don't want every label leaking into the catalogue. You decide when a label has earned it.

Phases

1. Scan DataStore.list(source_label). Abort unless

successes >= min_successes.

1. Compute the aggregate meta (backends seen, cameras seen, task names).
2. Generate SKILL.md (kind=task) from a template, filling in

description, backends, frontmatter tags derived from sampled metas.

1. Generate bundle.yaml with the default pipeline

(expert_replay → lerobot_build → pi0_finetune → evaluate_expert).

1. Write under roborsi/embodied/skills/<domain>/<new_task_name>/.

Success criteria

• SKILL.md + bundle.yaml land on disk.
• discover() after minting returns the new task.

Failure modes

Why

Without this skill, "auto-evolution" is aspirational. With it, the platform can point at its own footprint on disk and say "here are the N tasks I mastered by myself".