Echo framework for task status and action generation

Echo combines task status signal learning and action generation within a single Vision Language Action (VLA) model.

Echo is a framework that combines task status signal learning with action generation. It uses a single Vision Language Action (VLA) model. This model includes a compact design for status prediction.

The framework represents task status using stop and failure signals. These signals are added as two dimensions of each action chunk. They are denoised together with robot actions.

Echo learns these signals without frame-level progress annotations. It constructs successful, failed, and recovery trajectories. These trajectories have complementary supervision properties.

The framework controls interactions between signals and actions. It denoises the signals faster than the actions. This allows later action denoising steps to condition on cleaner status estimates.

Recognizing task completion and failure enables robots to stop upon success. It also allows them to initiate recovery or replanning after failure. Real-world manipulation experiments show Echo improves policy performance.

Key facts

  • Echo represents task status with stop and failure signals.
  • The framework learns signals without frame-level progress annotations.
  • Echo denoises status signals faster than robot actions.
  • Real-world manipulation experiments show Echo improves policy performance.

All sections

Edition