Skip to content

When will RECAP (π*₀.₆) training be supported? #857

Description

@toby018

Question

The recent π*₀.₆ announcement introduced RECAP, a reinforcement learning method that significantly improves performance over supervised learning. However, the current codebase only includes supervised learning implementations.

Are there plans to open-source the RECAP training code? If so, what is the expected timeline?

Missing Components

The current scripts/train.py implements supervised learning but lacks:

  • Value function training
  • Advantage computation
  • Advantage-conditioned policy architecture
  • Online RL training loop
  • Human correction integration

Context

RECAP achieves impressive results:

  • 2x+ throughput improvement on complex tasks
  • ~50% reduction in failure rates
  • Ability to learn from experience and corrections

This would be extremely valuable for the robotics community working on real-world deployments.

Request

Could the team provide:

  1. Timeline for RECAP code release (if planned)
  2. π₀.₆ base model checkpoint
  3. Guidance for community implementation (if official release is not immediate)

Thank you for the excellent work on OpenPI!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions