Echo: Coupled Action and Task Status Signal Denoising for Vision Language Action Models

Echo. Coupled Action and Task Status Signal Denoising for Vision Language Action Models. By Yufeng Liu and thirteen others, from Shanghai Jiao Tong University, Spirit AI, and four other institutions. Published on the twenty-second of October. Its code repository is worked out: about one hundred and ten source files, with a readme.

A robot that can tell when a task has succeeded can stop, and one that can tell when it has failed can start recovering or replan. Existing methods give vision language action policies, VLA policies for short, this ability by adding a prediction of task status signals that supports those execution decisions. Learning and wiring in those signals, though, can demand fine-grained supervision, separate monitoring models, or complicated architectures. And one possibility has been left largely unexplored: using the status signals to guide the actions themselves, rather than only to decide when to stop.

Echo answers both points within a single VLA. It learns task status without frame-level progress annotations, and it keeps the design for status prediction and action generation compact. Task status is expressed as two signals, stop and failure, and they are added as two extra dimensions of every action chunk. They are then denoised together with the robot's actions. Read together, the two signals tell three situations apart: execution still under way, successful completion, and failure.

The supervision question comes first. With no progress labels on individual frames to learn from, Echo constructs three kinds of trajectory instead: successful, failed, and recovery. Each kind supervises the signals in a different way, and the three are chosen because those properties complement one another. Status is learned from that combination.

To let status steer action, Echo does two things. It controls how the signals and the actions interact, and it denoises the signals faster than the actions. The status estimate therefore becomes clean earlier, and the later steps of action denoising can condition on that cleaner estimate, while the signal prediction itself stays accurate. Status is no longer just a readout at the end. It is an input the action generation can lean on as it goes.

The evaluation is extensive real-world manipulation. Across those experiments, Echo predicts task status accurately and consistently improves the policy's performance. Both halves of the design are tested at once: the signals have to be right, and the actions conditioned on them have to get better.

Knowing whether a task is done could mean dense labels or a second model watching the first. In Echo it is two more numbers in each action chunk, cleaned up ahead of the actions so that the robot's next moves can already take them into account.