VLDA learns compact deltas as robotic action interfaces
Original title: Vision-Language-Delta-Action: Learning Compact Deltas as Actionable Transition Interfaces for Robotic Manipulation
Implemented in LoongForge, it targets long-horizon consistency without predicting full future states.
Vision-Language-Delta-Action reformulates robotic manipulation as compact actionable transition modeling rather than direct observation-to-action mapping or full future-state prediction. A latent Delta variable sits between vision-language understanding and action generation, trained with Local Transition Filter and Global Transition Anchor. Delta-Driven Action Refinement grounds Delta in the action manifold and seeds action denoising. The method is implemented in the open-source LoongForge framework with two-stage optimization.
Unconfirmed: The paper does not report numeric gains for the claimed consistency and generalization improvements.
- LT-Filter keeps manipulation-centric changes and suppresses static reconstruction bias.
- GT-Anchor regularizes the overall direction of state changes.
- Authors report better long-horizon consistency, robustness, and generalization.
- Training uses LoongForge's modular distributed infrastructure.
From the original article
1 min read at huggingface.co
Robotic manipulation is not merely observation-to-action generation, but the realization of task-specified state changes in the physical world. Existing vision-language-action (VLA) methods directly learn observation-to-action mappings, imposing a substantial learning burden due to the gap between semantic understanding and action execution.
By Lingsen Zhang, Wei Li, Jie He · 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.