Update patched_factory.py

Update patched_factory.py

Fix: Adaptive gripper threshold with debug logging

Force output_hidden_states=True in lang_encoder forward call

Debug: Check if hidden_states are actually enabled

Final fix: Exact checkpoint dimensions (action=6, not 7)

Fix: Use correct LSTM dimensions (4 layers, hidden=1024)

Fix: Map checkpoint policy weights correctly and enable hidden states

Add LSTM policy head for robot action prediction

Cache checkpoint in /data + diagnose missing policy head

Add debug output to see model output format

Remove simulation fallback - real model only

Fix: Convert attention_mask to boolean before passing to model

Patch OpenFlamingo to remove 'labels' before calling MosaicGPT

Fix: Remove 'labels' argument from model forward pass

Use patched factory to bypass BCFlamingo init bug

Add einops-exts==0.0.4 dependency

Remove large eval files

Deploy real interactive model

Load checkpoint directly in app.py

Load trained RoboFlamingo checkpoint from HuggingFace (robovlms/RoboFlamingo)

Add custom factory wrapper to avoid BCFlamingo argument issues

Remove problematic arguments from create_model_and_transforms

Add zarr and RoboFlamingo dependencies

Use robot_flamingo.models.factory for REAL RoboFlamingo

Fix: Use actual RoboFlamingo structure

Deploy REAL RoboFlamingo with LSTM policy head and trained checkpoint

Complete deployment: Real interactive RoboFlamingo model with correct shapes

Complete deployment: Real interactive RoboFlamingo model with correct shapes

Remove decoder_type parameter for vanilla OpenFlamingo

Add einops-exts dependency

Fix imports: use open_flamingo.src.factory

Deploy with src only (no large eval data)

Deploy real interactive RoboFlamingo model

Add pre-computed results viewer

Initial commit: RoboFlamingo demo with fixed dependencies

Update requirements.txt

Create requirements.txt

Create app.py

initial commit