Event
VLDA proposes modeling robot state transitions with compact Delta
1 reports · 1 sources
Researchers propose the Vision-Language-Delta-Action (VLDA) paradigm, which uses compact Delta variables to model task-relevant state transitions, replacing VLA's direct observation-to-action mapping and world-action models' prediction of complete future states. It uniformly learns the target Delta through LT-Filter and GT-Anchor, and then DDAR injects the Delta into the action denoising process as a structured seed. This method is implemented in the open-source training framework LoongForge, and the authors state that experiments show VLDA improves long-horizon consistency, robustness, and generalization ability.
Report timeline
- Vision-Language-Delta-Action: Using Compact Deltas as an Executable State-Transition Interface for Robotic Manipulation
VLDA proposes modeling task-relevant state transitions with compact Delta variables, replacing VLA's direct observation-to-action mapping and the full future-state prediction of world-action models. It learns Delta under a unified objective via LT-Filter and GT-Anchor, and DDAR injects Delta into the action denoising process as a structured seed. Experiments implemented in the open-source training framework LoongForge show that VLDA improves long-horizon consistency, robustness, and generalization.
HuggingFace Daily Papers ·