vla_real_pk_remove_sharp_gradgate_step400

Edited pi0.5 VLA checkpoint for pass_knife task — pk_remove_sharp_gradgate arm at step 400.

Deployment goal

Remove unsafe 'sharp' (no-rotation) handoff behavior.

Edit recipe

• steering_mode: hidden_v9_mc_softhybrid_precommit_gated
• target_subset: {0:left, 1:right}
• loss_formula: -logsumexp(z_target_subset) + logsumexp(z_others)
• gating: GRADIENT GATE — top-10% of unpref frames by |∂(v9_loss)/∂h|, computed offline
• gate_calibration: 296 of 2951 sharp-episode frames (~10% of trajectory mass), threshold |grad|>=32.29
• ablation_arm: GRAD-GATE (experimental)

Common hyperparameters

γ=0.1, β=1.0, lr=1e-5, batch=32, num-steps=400-600, save-interval=50-100, ViT frozen

Foundation VLA

pi05_real_pk_mixed/real_pk_mixed_v3 step 24999 — frozen mixed-mode foundation, edited only on the action-expert + LLM (ViT frozen).

Classifier used

/mnt/data3/classifiers/real_v3/pk_v5h_mc/best_v5h_mc.pt (v5h-mc 3-class softmax, val_acc 91-94%)

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

pk_remove_sharp_full — 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_pk_mixed