vla_real_pb_from_left_edit_step400

Edited pi0.5 VLA checkpoint for push_block task — pb_from_left arm at step 400.

Deployment goal

Bias model toward 'from_left' approach direction (when scene/policy permits both).

Edit recipe

• steering_mode: action_v4 (action-space classifier-guided)
• classifier: v3-style (hidden + action + progress → binary)
• loss_formula: γ * (-log P(target | h, a, p)) gradient on action chunk
• gating: frame_index < 52 (~32% retention, sim 1pillar locked recipe semantic)
• ablation_arm: main (matches 1pillar locked recipe with bs=32 instead of 8)

Common hyperparameters

γ=0.1, β=1.0, lr=5e-5, batch=32 (was 8 in 1pillar locked), num-steps=400, save-interval=100, ViT frozen, --classifier-action-dim 8

Foundation VLA

pi05_real_pb_mixed/real_pb_mixed_v3 step 24999 — frozen mixed-mode foundation, edited only on the action-expert + LLM (ViT frozen).

Classifier used

/mnt/data3/classifiers/real_v3/pb/classifier_from_{left,right}/best_classifier.jax.pkl (v3-style binary, hidden+action+progress)

Why this checkpoint?

Top-2 by composite val_loss_pref + 0.1 * loss_redirect. The composite score balances target-mode preservation (val_loss_pref low) and active editing pressure (loss_redirect strongly negative). For all 4 main edits, the LATEST ckpts won by composite score — i.e., the editing benefits accumulate throughout training and don't overfit before the final step in this configuration.

Ablation companion

pb_from_right (mirror, target=from_right) — eval both and compare to validate the gating choice.

Eval target

50-seed real-robot rollouts. Compare:

• Target rate: fraction of episodes where deployment goal is achieved
• Overall SR: model still completes the task successfully
• vs foundation VLA baseline (no editing) on the same seeds

Loading

from openpi.training import config as _config
import openpi.shared.array_typing as at
from openpi.models.model import preprocess_observation
# Load this repo's params/ subdirectory as the model checkpoint.
# Use config: pi05_real_pb_mixed