Downloads · 30 days
8
9% of all-time downloads
Ashenzari/vla-lstm-success-monitor
vla-lstm-success-monitor is a robotics model from Ashenzari. Use it for the robotics task on the model card, and read the license before you ship it in a product. It is set up for pytorch.
This revision contains a five-fold LSTM ensemble trained to estimate causal robot-task progress from frozen VLA/VLM backbone features. It is the progress-only successor to the v1 success/failure monitor.
Downloads · 30 days
8
9% of all-time downloads
All-time downloads
86
Public
Repo size
232 MB
Likes
0
Public
Click a slice to open those files.
.pt95.1 MB · 100%
From the Hugging Face model README
This revision contains a five-fold LSTM ensemble trained to estimate causal robot-task progress from frozen VLA/VLM backbone features. It is the progress-only successor to the v1 success/failure monitor.
v2.1.0progress-onlycoffeecup, saucer, waterbottle, winenew_vla_checkpoint_vlm_30kCached features from v2.0.0 are not compatible with these temporal heads.
Re-extract features with the 30k-step VLA backbone before inference.
The model architecture still contains an outcome head for checkpoint
compatibility, but that head was not trained for this release. Do not interpret
its logits as success/failure probabilities. Only head_progress output is valid.
Install the source repository and PyTorch when loading this LSTM release. The rank normalizer stored in each checkpoint must be applied before inference.
from huggingface_hub import snapshot_download
import json
import torch
from monitor.feature_normalization import TaskRankNormalizer
from monitor.model import Monitor
root = snapshot_download("Ashenzari/vla-lstm-success-monitor", revision="v2.1.0")
config = json.load(open(f"{root}/config.json"))
entry = config["checkpoints"][0]
checkpoint = torch.load(
f"{root}/{entry['file']}", map_location="cpu", weights_only=True
)
model = Monitor(temporal="lstm", dropout=checkpoint["args"]["dropout"])
model.load_state_dict(checkpoint["model"], strict=True)
model.eval()
normalizer = TaskRankNormalizer.from_state_dict(checkpoint["rank_normalizer"])
# pooled_features: [T, 4096]
x = normalizer.transform(pooled_features.float(), task_id="wine")
with torch.inference_mode():
_, progress = model(x.unsqueeze(0))
progress = progress[0] # [T], range [0, 1]
For ensemble inference, repeat the same operation for all five checkpoint entries and average their progress curves.
Features were extracted from the new_vla_checkpoint_vlm_30k VLA checkpoint.
Each prediction step uses the context offsets [-0.3, -0.2, -0.1, 0.0]
seconds, 64 visual tokens, mean pooling, and a 4096-dimensional feature vector.
Changing the VLA/VLM checkpoint can shift the feature distribution and requires
revalidation.
See EVALUATION.md and config.json for the clean and freeze/reverse robustness
metrics. The evaluation is in-domain and does not establish zero-shot
performance on unseen tasks.