🚨 MoNaVLA Perception Audit: Critical Failure Report
⚠️ CRITICAL ERROR: Vision-Language Alignment Mismatch Detected!
Backbone model is unable to process visual grounding. Repeating <patch_index_1024> tokens.
Backbone model is unable to process visual grounding. Repeating <patch_index_1024> tokens.
BACKBONE: BLIND
(Repeat loop: Vision Mismatch Detected)
frame_001.jpg
VLM Raw Output:
Action Prob (F): 0.218
Action Prob (FL): 0.202
BACKBONE: BLIND
(Vision Mismatch)
frame_005.jpg
VLM Raw Output:
Action Prob (F): 0.308
Action Prob (FL): 0.210
BACKBONE: BLIND
(Perception Failure)
frame_010.jpg
VLM Raw Output:
Action Prob (F): 0.316
Action Prob (FL): 0.208
👨🏫 Professor's Briefing Note:
1. Visual Blindness (VLM Error): The backbone fails to generate grounding coordinates. It is currently "visually blind" to the basket.
2. Policy Bias: Even with logit penalty, the model defaults to FORWARD (Idx 1) because the visual input is too noisy/broken to trigger turning logic.
3. Conclusion: Urgent LoRA Re-alignment needed. The Vision-Tower outputs are not reaching the Policy Head correctly.