Skip to content

[Feature Request] Automatic Data Generation Pipeline #110

Description

@Papaercold

Motivation

Following the discussion with Zeyu, I propose enabling automatic data generation for LeIsaac. This will allow us to generate task executions, trajectories, and structured datasets directly from simulation, rather than relying solely on manual demonstrations.

Implementation Plan

I will start migrating the core auto-generation logic from so101-autogen to LeIsaac.

Key Goals:

  • Enable large-scale synthetic dataset generation.
  • Support Hybrid IL + RL pipelines.
  • Accelerate iteration on new tasks.

Status: I am starting the code migration now.

Activity

  1. Papaercold commented on Jan 28, 2026

    @Papaercold
    ContributorAuthor

    I’m basically done. Right now I’m doing the final tweaks and refactoring of the data generation state machine. I’ll open a PR once it’s finished.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions