real_robot_v4action_classifiers

v3-style HiddenStateClassifier checkpoints (hidden + action + progress -> P(target)) for real-robot pk (Pass Knife) and pb (Push Block) tasks. Used as the cg_loss target in action_v4 mode editing on pi0.5 VLA, and as a sidecar inference logger via scripts/serve_policy_with_classifier.py (in openpi).

All classifiers trained with the v4action defaults (lam_grad=1.0, lam_corr=2.0, grad_min_thresh=1.5) — substantially more action-aware than the original v3 classifiers. See docs/real_robot_v2.md §9.8–9.10 in behavior-uncloning for full probe results.

Files

file	task	semantics	a=0 ablation acc	grad ratio
pk_left.pt	pk	binary: 1 if mode==left, else 0	24.3%	0.62
pk_right.pt	pk	binary: 1 if mode==right, else 0	0.2%	0.70
pk_sharp.pt	pk	binary: 1 if mode==sharp, else 0	8.2%	0.65
pk_keepLR.pt	pk	binary: 1 if mode in {left,right}, else 0 (synthetic)	—	—
pb_from_left.pt	pb	binary: 1 if mode==from_left, else 0	15.4%	0.50
pb_from_right.pt	pb	binary: 1 if mode==from_right, else 0	75.2%	0.53

(a=0 ablation acc and grad ratio are measures of action-sensitivity. Lower a=0 acc and higher grad ratio = stronger steering signal.)

Architecture (mirrors train_hidden_classifier_v4.py:V3Classifier)

hidden (2048-D)  → hidden_proj (2048→256, LayerNorm, ReLU)        ─┐
action (8d × 10 steps = 80-D)                                       ├─ concat (400-D)
progress (sinusoidal embed, 64-D)                                   ─┘
  → backbone (400→512, LayerNorm, ReLU, →256, LayerNorm, ReLU)
  → dropout(0.5) (eval=identity)
  → head (256 → 1) → sigmoid → P(target | h, a, p)

All checkpoints are PyTorch state_dicts. Load with:

import torch
from openpi_serve_policy_with_classifier_helper import HiddenStateClassifier
sd = torch.load("pk_left.pt", map_location="cpu")
clf = HiddenStateClassifier(action_dim=80)
clf.load_state_dict(sd); clf.eval()

(or use scripts/serve_policy_with_classifier.py directly which loads them automatically)

Usage in serve_policy_with_classifier.py

huggingface-cli download haohw/real_robot_v4action_classifiers \
  --local-dir ~/v4action_classifiers

uv run scripts/serve_policy_with_classifier.py \
  policy:checkpoint --policy.config pi05_real_pk_mixed --policy.dir <edited-ckpt> \
  --classifier-name=left  --classifier-ckpt=$HOME/v4action_classifiers/pk_left.pt \
  --classifier-name=right --classifier-ckpt=$HOME/v4action_classifiers/pk_right.pt \
  --classifier-name=sharp --classifier-ckpt=$HOME/v4action_classifiers/pk_sharp.pt

Each inference call will print one log line:

[classifier @ infer #42] P(left)=0.034 | P(right)=0.812 | P(sharp)=0.154
    argmax=right  margin=0.658  infer_ms=187.3