Joeyfully/smolvla_rlt_libero_10

SmolVLA-RLT LIBERO-10

This repository contains an offline / batched RLT-style residual adaptation for SmolVLA on LIBERO-10.

Released artifacts

• Trained residual actor checkpoint
• Trained action-latent autoencoder checkpoint
• Exported SmolVLA-RLT policies
• Qualitative case-study videos

Important notes

This is not a full online RLT reproduction. It is an offline residual adaptation on top of a frozen SmolVLA policy. The frozen SmolVLA base policy is required separately. Before use, update base_policy_path in the exported policy config to point to your local SmolVLA checkpoint.

Qualitative observations

We observed several paired rollout patterns:

1. Base fails, RLT succeeds.
2. Base succeeds, RLT succeeds more smoothly.
3. Both fail, but RLT makes more task progress.
4. Tug-of-war failure: the base policy chooses the wrong object or order, and the residual branch can only locally oppose it.
5. Larger residual scales may over-correct precise contact tasks.

Reproducibility

This release supports checkpoint-level and qualitative reproducibility. Full training reproduction requires the same local SmolVLA base policy, LeRobot/LIBERO setup, and derived RLT token data.