Overview
DELE-w0.5 represents a departure from conventional World-Action Models (WAMs) that typically integrate video-generation backbones for robot control. The core design principle of DELE-w0.5 is to infer robot actions directly from predicted future latent states, which encapsulate the physical outcome relevant to an action, rather than from a dense sequence of visual frames. This methodology posits that generating intermediate visual trajectories is an 'unnecessary intermediate objective' for robotic manipulation, consuming substantial model capacity and computation without directly specifying the intended physical outcome.
The model’s approach focuses on predicting the state the world will achieve after an action, decoupling this from the visual appearance evolution frame by frame. This formulation aims to remove the 'high-dimensional visual redundancy' associated with dense video representations, which is suggested to facilitate 'cheaper training' and 'low-latency inference'.
Research Context
Traditional World-Action Models (WAMs) construct robot control systems utilizing video-generation architectures. These systems are designed to jointly predict both dense future visual trajectories and the corresponding robot actions. The underlying assumption in such models is that a comprehensive visual reproduction of how the world appears at every intermediate moment is essential for effective robot control. However, the researchers behind DELE-w0.5 contend that for robotic manipulation tasks, the primary goal of a world model is to ascertain the physical state the world will reach post-action execution. They suggest that the visual transitions observed in intermediate frames primarily describe the visual evolution between physical states without directly conveying the physical outcome intended by the robot's action.
Approach
DELE-w0.5 was designed to infer robot action sequences based on their corresponding 'compact future latent state'. This future latent state functions as an explicit bridge, linking world modeling processes with action generation. It is engineered to capture the 'action-relevant physical outcome' of robot interaction. By focusing on how the physical world transforms under robotic actions, instead of its visual appearance changes frame by frame, DELE-w0.5 aims to circumvent the inefficiencies associated with dense video representations. This strategic focus is intended to eliminate high-dimensional visual redundancy, thereby enabling more efficient training and faster inference times.
Findings
In a series of real-robot trials, DELE-w0.5 demonstrated superior performance compared to existing policies. The model was evaluated across 480 real-robot trials involving four distinct 'long-horizon manipulation tasks'.
- DELE-w0.5 achieved an 'overall full-task success' rate of 62.5%.
- The model exhibited an '81.3 macro ordered-stage progress'.
- These results indicate that DELE-w0.5 outperformed the 'strongest baseline' by significant margins, specifically by 47.5 percentage points in 'overall full-task success' and 30.7 percentage points in 'macro ordered-stage progress'.