---
source: arxiv
url: http://arxiv.org/abs/2604.05614v1
published_at: '2026-04-07T09:03:12'
authors:
- Theodor Wulff
- Federico Tavella
- Rahul Singh Maharjan
- Manith Adikari
- Angelo Cangelosi
topics:
- vision-language-action
- hierarchical-policy
- language-action-alignment
- preference-learning
- robot-transparency
relevance_score: 0.91
run_id: materialize-outputs
language_code: en
---

# Grounding Hierarchical Vision-Language-Action Models Through Explicit Language-Action Alignment

## Summary
GPLA trains a hierarchical vision-language-action model to make its generated subtask language match the robot's visual scene and action trajectory more closely. It does this with a learned grounding scorer and preference tuning, aiming to reduce the need for manual subtask annotations.

## Problem
- Hierarchical VLA models can output both language substeps and robot actions, but their training usually does not force those two outputs to agree with each other.
- This matters for robot transparency and human-robot collaboration: a robot explanation is only useful if the described subtask matches what the robot sees and does.
- Existing evaluation often focuses on task success or action accuracy, while missing whether intermediate language is actually grounded in observation and behavior.

## Approach
- The paper builds a hierarchical VLA with a high-level Gemma-3-4B-IT VLM that generates low-level instructions and a low-level SmolVLA that predicts 8-step action trajectories on LanguageTable.
- It trains a separate grounding model that maps vision-action pairs and text into a shared embedding space. The model uses frozen SigLIP 2 vision/text encoders, a small transformer for actions, and FiLM layers to condition vision features on actions.
- The grounding model is trained with symmetric InfoNCE contrastive loss plus a diversity regularizer, so matched language and vision-action pairs score higher than mismatched pairs.
- During GPLA training, the hierarchical VLA samples multiple language-action candidates for the same input, the grounding model ranks them, and the highest-scoring and lowest-scoring outputs form chosen/rejected preference pairs.
- The high-level VLM is then updated with SimPO preference optimization so it generates subtask descriptions that better align with the resulting actions and scene, without requiring new manual preference labels.

## Results
- On LanguageTable, supervised training gives the best text-overlap metrics for low-level instruction generation: BLEU **0.111 ± 0.05**, ROUGE **0.405 ± 0.12**, METEOR **0.313 ± 0.12**, BERTScore **0.984 ± 0.00**.
- GPLA variants score lower on token-overlap text metrics, for example GPLA with the action-conditioned grounding model reaches BLEU **0.063 ± 0.05**, ROUGE **0.300 ± 0.12**, METEOR **0.218 ± 0.12**, BERTScore **0.980 ± 0.00**.
- Action quality stays close to the supervised baseline. Supervised has MSE **0.046 ± 0.02** and MAE **0.164 ± 0.04**, while GPLA (action-conditioned) has MSE **0.045 ± 0.02** and MAE **0.163 ± 0.04**.
- GPLA therefore claims action performance comparable to fully supervised fine-tuning while using preference pairs generated by a learned grounding score instead of costly extra annotations.
- Among GPLA variants, CLIP, SigLIP 2, and the action-conditioned grounding model produce very similar trajectory metrics; the paper's main claim is the explicit language-action grounding mechanism rather than a large raw performance jump.
- The paper also claims the action-conditioned grounding model yields a better shared embedding space for language and vision-action alignment than off-the-shelf vision-language encoders, based on embedding-space analysis and qualitative examples.

## Link
- [http://arxiv.org/abs/2604.05614v1](http://arxiv.org/abs/2604.05614v1)