SOUL unlearning reduces state hallucination in robot VLA policies
Original title: Sparse Feature Policy Unlearning Mitigates State Hallucination in Vision-Language-Action Models
The method forgets hallucination-linked sparse features while retaining successful manipulation behavior in simulation and real robots.
Researchers describe state hallucination, where vision-language-action models keep acting as if an unrealized robot-object state had occurred. They link the failures to weaker attention on task-relevant visual regions and to sparse features found with autoencoders. Their SOUL method selectively unlearns those policy features and reports fewer failures and higher task success without a large loss of manipulation ability.
Unconfirmed: Reported gains come from the authors' experiments; independent replication is not stated.
- Hallucination coincides with weakened attention to task-relevant visual regions.
- Sparse features from failures and successes are forgetting and retention targets.
- Tests span multiple VLA architectures in simulation and real environments.
- Authors say existing manipulation capabilities are not substantially compromised.
From the original article
1 min read at huggingface.co
Vision-Language-Action (VLA) models have shown strong generalization in robotic manipulation by leveraging rich representations from pretrained vision-language models. However, their deployment in real-world environments remains limited by recurring unreliable behaviors.
By Jiho Lee, Jeongeun Park, Heayoun Choi · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.