CycleVLA connects two recovery decisions: which subtask should the robot revisit, and which predicted action chunk should it execute on the retry? A vision-language model selects the earlier subtask when a prerequisite is missing. The action policy then samples several chunks from a fresh observation, and minimum Bayes risk (MBR) decoding selects one representative candidate.
The connection matters because returning to an earlier robot configuration does not recreate the earlier scene. A dropped object can now lie somewhere else. The retry must respond to that current scene, then choose an action sequence suited to it. The CycleVLA project presents this as a cycle of progress-aware inspection, subtask backtracking and renewed execution.
Progress invites inspection; stop confirms a transition
Section IV-A of the paper extends a seven-component robot action with two scalar outputs. The original components describe three translation deltas, three rotation deltas and the gripper signal. The additions are a stop signal and a progress signal, producing nine components in total.
These signals have different jobs. Progress indicates that execution is approaching the end of the current subtask, so it can trigger a VLM inspection. Stop identifies termination precisely enough to advance to the next subtask. In the paper’s inference algorithm, an InProgress phase handles the inspection decision; a Complete phase continues execution until the stop signal confirms termination.
The progress training target comes from the normalized timestep within a labeled subtask, discretized in increments of 0.1. It therefore encodes a position within that demonstrated subtask rather than a calibrated probability that the task has succeeded. Stop is a separate binary target, with the final subtask action oversampled during training. Both signals are predicted alongside the action values, rather than through additional classification heads.
The subtask labels also have a concrete origin. The authors ask an LLM to decompose the task instruction. In parallel, they extract gripper-state segments and movement primitives from robot state differences. When the proposed subtask count matches the gripper-segment count, they pair the sequences directly. Otherwise, an LLM uses the movement-primitive sequence to infer the timestamp boundaries. This creates the temporal annotations needed to train progress and stop together.
The VLM selects a missing prerequisite to revisit
When predicted progress reaches the inspection threshold, the method provides the VLM with synchronized third-person and wrist-camera views, the current subtask, and the full subtask list. The views serve complementary roles: the third-person view supplies scene and object context, while the wrist view provides closer cues about alignment and contact.
The VLM chooses either to proceed or to backtrack to the earliest subtask that restores a missing precondition. The paper gives a dropped grasped object as an example: the recovery target is the grasping subtask. This makes the return point part of the decision, rather than automatically repeating whichever action happened most recently.
In a hypothetical pick-and-place sequence, the robot might be near the end of transport with an empty gripper. Repeating the final placement motion would start from the same missing prerequisite. Re-entering the grasping subtask gives the action policy a task whose objective is to obtain that prerequisite again.
Once the VLM permits normal progression, the separate stop signal still controls the precise subtask transition. This keeps “inspect whether the subtask is on track” distinct from “the current subtask has ended.”
Backtracking returns the robot, then the retry observes the scene
Section IV-B describes backtracking through reverse execution of recorded robot delta actions. Algorithm 1 also clears the queued actions and obtains a new observation after this return. The policy then samples retry candidates conditioned on that observation and the selected earlier subtask.
The object’s state is carried forward. In the hypothetical dropped-object example, the robot returns along its recorded motion, but the object remains at its current location. The next grasp therefore depends on perceiving where it is now. Reusing the old action queue would instead retain commands selected for the earlier scene, which is why discarding the queue and observing again are meaningful parts of the recovery path.
This is the specific mechanism behind the method’s cyclic execution: revisit a subtask, obtain the current evidence, and generate a new attempt. The broader long-horizon task memory and recovery guide explains how completed effects and recovery state can be retained around such a loop.
MBR picks a sampled trajectory with the smallest average distance
For the retry, the VLA draws N action chunks from its stochastic policy using the same current observation and subtask. Each chunk contains H future steps. The paper accumulates the predicted translation and rotation deltas into a trajectory represented by 6H pose features, then computes distances between those trajectories.
For each candidate, MBR averages its distance to all N candidates, including itself. The candidate with the smallest average distance is selected. This is a medoid-like choice: the output is one of the sampled chunks. It keeps the commands of that chunk intact rather than constructing a new command sequence by averaging the samples.
Consider the following hypothetical distance matrix, in arbitrary distance units. These are illustrative candidate distances, not measured CycleVLA results.
| Candidate | Distance to A | Distance to B | Distance to C | Average distance |
|---|---|---|---|---|
| A | 0 | 2 | 8 | 10/3 |
| B | 2 | 0 | 6 | 8/3 |
| C | 8 | 6 | 0 | 14/3 |
B has the smallest average, so the decoder selects B’s original action chunk. The calculation compares the sampled trajectories with each other; it does not receive a ground-truth success score for each candidate. Its preference is for a representative trajectory under the policy’s sampled distribution.
The accumulated trajectory representation also explains why this comparison extends beyond the next action. Two chunks can begin with similar deltas and diverge later. Evaluating their multi-step pose features includes those later differences in the distance used for selection. The selected chunk is then placed into the action queue for execution.
Where the two decisions meet
Backtracking and MBR operate at different resolutions. The VLM changes the subtask that the robot should attempt, using the scene and missing preconditions. MBR chooses among proposed motions for that selected subtask, using agreement between sampled trajectories. The observation taken after backtracking joins them: it is the input from which those retry motions are generated.
This reading follows the current paper’s method and inference algorithm; no model or robot was run for this article. It leaves a concrete way to inspect a retry: identify the selected subtask, the fresh scene it received, the candidate trajectories and the one candidate chosen by the distance rule. Those values explain the recovery computation more directly than a single label saying that the robot tried again.