Echo Teaches VLA Robots to Recognize Success and Failure Inside Their Own Actions

A robot that picks up a cup and puts it on a shelf needs to do more than move its arm. It has to know when to stop, and it has to recognize that the cup has fallen and it is time to change course. A team of 14 researchers (Yufeng Liu, Bohao Fan, Juntu Zhao, Cheng Ma, Yang Gao, and others) has published Echo. It is a way to build this awareness directly into a vision-language-action model: two status signals, stop and failure, become extra dimensions of every action chunk and are processed together with the motions.

Status as part of the action itself

VLA models generate actions in chunks: a single pass produces a short sequence of commands for the manipulator. Echo adds two numbers to each chunk. One says "the task is done, time to stop"; the other signals "something has gone wrong." Together they yield three states: the task is still in progress, it has finished successfully, or it has failed.

Previously, such decisions usually required one of three things: detailed progress labeling, an external observer model, or a complex architecture with separate branches. Echo's authors avoid all three. Everything lives inside a single VLA, and the signals are denoised by the same mechanism as the robot's motions.

How to avoid per-frame labeling

To teach the model to recognize status without labeling every frame, Echo collects three kinds of trajectories whose labeling properties complement one another:

• successful trajectories: show what a correct completion looks like and when a stop is needed;
• failed trajectories: provide examples of failure, from which the failure signal learns;
• recovery trajectories: show how the robot gets out of an error state and continues working.

These data types back each other up: where one kind gives weak supervision, another strengthens it. Episode-level or segment-level labels are therefore enough, and there is no need for a progress estimate on every frame.

Signals are denoised faster than motions

Echo's most unusual design choice concerns the order of denoising. Diffusion-like or flow-like action generation proceeds in steps, from noise to a clean result. The authors control the interaction between signals and actions and force the status signals to converge faster than the motions themselves.

As a result, in the late steps of generation the actions rely on an almost clean status estimate. If the model has "understood" that the task is wrapping up, that knowledge influences how the final commands are formed. The signal prediction itself stays accurate: the controlled interaction keeps the motions from contaminating the status estimate.

According to the authors, the experiments were run on real manipulation tasks. Echo predicts task status accurately and consistently improves policy results. Per-task numbers and comparisons with baseline methods are given in the paper itself (arXiv 2610.31569), and the code is published in the Echo-Sim repository.

What developers get

For those building manipulation systems on VLA models, the approach has several practical advantages:

• no separate observer model is needed: the same network that predicts actions predicts status;
• no per-frame progress labeling is needed, which noticeably lowers data collection costs;
• stopping on the stop signal eliminates unnecessary motions after success;
• the failure signal can serve as a trigger for a retry or for replanning;
• architectural changes are minimal: two extra dimensions in the action chunk and a tuned denoising schedule.

Because the signals are simply appended to the policy output, the idea can be tested on existing VLA models with a generative action head. This requires a set of demonstrations that includes successes, failures, and recoveries. Such data is expensive to collect, but it is cheaper than per-frame annotation. It remains to be seen how the approach scales to longer multistep tasks and to different robots. The project is open: the GitHub Pages site and the repository make it possible to reproduce the idea and compare it with setups where an external module estimates status.