swoosh-data/lego-pi05-lora-h50-6812

LEGO π0.5 LoRA H50 — Step 6,812

LoRA adapter for the LeRobot pi05_base policy, trained on bimanual LEGO assembly demonstrations with an action horizon of 50.

Contents

• —Final LoRA weights and PEFT configuration.
• —Preprocessor, postprocessor, and normalization assets required for inference.
• —Complete training configuration.
• —Deterministic 435-sample held-out evaluation and evaluator smoke test.

Optimizer state and intermediate checkpoints are intentionally omitted because this is an inference/evaluation archive rather than a resumable training bundle.

Training

• —Final checkpoint: step 6,812
• —Base model: lerobot/pi05_base
• —Training mode: LoRA/PEFT, rank 16
• —Action horizon: 50
• —Relative joint actions; grippers remain absolute
• —Dataset: swoosh-data/lego_assemblies_labeled_pi05
• —W&B run ID: w3heafk6

Held-out evaluation

Evaluation used 435 deterministic semantic-span samples, four fixed flow-matching draws, and an execution horizon of five.

Complete metrics and per-phase/per-dimension breakdowns are under evaluations/.