FPC-VLA calls its vision-language supervisor when the gripper command changes enough to cross a threshold. If the supervisor rejects the proposed action, the published algorithm adds a translation and yaw correction to the fused action. It does not directly change the gripper command, roll or pitch through that correction vector. Those choices define both the opportunity for correction and the failures this particular supervisor can miss.
The FPC-VLA project presents a VLA actor, an action-fusion module and a separate supervisor. The details below follow paper version 2, especially Sections 3.3 and 3.5. This is a reading of the method, checked October 15, 2026; no policy, supervisor or robot was executed for this article.
A keyframe is a command transition, not an arbitrary video frame
Algorithm 2 receives a seven-component action from the fusion module: three positional increments, three rotational increments and a gripper value. It compares the current gripper command with the preceding one. The supervisor is queried when the absolute difference exceeds the threshold δg.
This makes the intended opening or closing transition the decision point. The algorithm receives the current image and asks the VLM whether the action is appropriate. If the threshold is not crossed, it executes the fused action without a supervisor query. A response that approves the decision also leaves the fused action unchanged.
The distinction between command and physical state matters. A new close command does not, by itself, establish that the fingers reached the object or formed a stable grasp. Likewise, an object can begin slipping while the command continues to request the same closed state. Neither event necessarily creates the command transition used by this trigger.
For example, imagine a gripper approaching a cup. A proposed close command can trigger a review of whether the gripper is positioned to grasp it. A later slip during transport is a different event: the command may remain closed throughout.
“No” produces a restricted action adjustment
In Section 3.5, a response beginning with “No” supplies structured directional corrections. Direction tokens and small or large magnitude indicators map to predefined step sizes. The algorithm extracts three translation corrections and one yaw correction, then adds them to the fused action.
The correction vector has the form [Δx, Δy, Δz, 0, 0, Δrz, 0]. Read the zeros as part of the method:
| Action component | Effect of the supervisor’s additive correction |
|---|---|
| Translation along x, y and z | Can be adjusted by the parsed directional correction |
| Yaw, represented by the z rotation increment | Can be adjusted |
| Roll and pitch increments | Retain their values from action fusion |
| Gripper command | Retains its value from action fusion |
The paper defines positional displacement in the robot’s base frame. Its parsed directions use that base-frame convention and the method’s predefined magnitude mapping. An adapter must map the direction token and step size into those coordinates.
There is also a useful reading detail in the equations. The prose describes selectively overriding corrected components, while Equation 19 and Algorithm 2 express an addition to the fused action. The concrete operation shown is an additive adjustment. It is not a replacement of the entire action by a new seven-component command from the supervisor.
Consequently, rejecting a grasp decision does not mean that this correction vector cancels the gripper transition. It changes the listed motion components while retaining the incoming gripper value.
Action fusion is separate from supervisor correction
Before supervision, FPC-VLA combines predictions for the same execution timestep made from different observation histories. These are not simply the last commands that the robot already executed. Each historical prediction is split into a six-component pose increment and a scalar gripper value.
For pose fusion, the method compares each historical pose prediction with the most recent prediction using cosine similarity. It combines the transformed similarity with temporal decay and normalizes the weights. Predictions more aligned with the current pose prediction, and more recent predictions, influence that fused pose through this weighting rule.
The gripper stream uses normalized temporal-decay weights without the pose-similarity term. The authors separate it because pose similarity need not describe gripper intent: the arm can move while the gripper continues holding an object. The supervisor then receives the fused action and checks the gripper-change condition described above.
This separation also limits an easy misreading of the correction table. A zero in the supervisor’s gripper correction means “do not add a gripper adjustment here.” It does not mean that the fusion module never changes a gripper command or that the physical gripper remains fixed.
The trigger leaves a specific blind spot
The paper’s limitations section explicitly identifies post-grasp slipping and dropping as problems the gripper-change-only trigger cannot address. The authors also note the visual-input limitation for perceiving contact forces and the limited complexity and duration of their long-horizon evaluations.
For a reader adapting this architecture, the first concrete question is therefore whether the intended failure produces a gripper-command transition at all. If it does not, changing the supervisor prompt alone leaves the trigger unchanged. If it does, the next question is whether translation and yaw are sufficient correction dimensions for that failure.
Use the VLA evaluation guide for whole-system success, latency and intervention measurement. FPC-VLA adds a narrower mechanism to understand first: when the supervisor is called, how its answer changes the fused action, and which components it leaves untouched.