A new preprint on arXiv, Replan for Error Recovery: Training Embodied VLMs through Verifiable Interaction with Action-Grouped Turn-GSPO, addresses a practical gap in embodied vision-language models (VLMs). These models can carry out increasingly complex manipulation tasks, but when execution goes wrong, the resulting environment state may demand replanning rather than simple continuation.
The paper notes that current recovery methods tend to depend on inference-time mechanisms. The proposed alternative is to train the model through verifiable interaction, using an approach called action-grouped turn-GSPO. This shifts the focus from patching errors at inference time to building recovery behaviour into the training process itself.
Because the abstract is brief, the details of the action-grouped turn-GSPO method are not fully specified here. The contribution, as described, is a training-oriented route to error recovery for embodied VLMs, rather than a purely reactive strategy.