Self-Supervised Critical Phase Detection for VLA Refinement
Abstract
Vision-Language-Action (VLA) policies handle most of a manipulation rollout, then fail on the few decisions that actually matter — the instant a peg first tilts inside a hole, the moment a grasp commits. We call these moments critical phases and define them by decision sensitivity: how much a small change to the action at that timestep flips the eventual success or failure of the rollout.
Crucially, a critical phase is not the same as a likely failure. Successful rollouts pass through the very same decision points, so criticality is a property of the decision structure, not of failure probability. This separates our target from runtime failure detection, which measures how out-of-distribution the current step looks and therefore stays quiet on successes and reacts only after a failure has surfaced.
We study how to detect critical phases without failure labels and without a per-task success oracle, by combining two signals of very different character. A force/torque (F/T) sensor measures contact force and torque directly, so it catches the moments contact begins or breaks at far lower latency than inferring them from camera images — but it has no idea which task, or which part of one, is underway. The VLA's generalization reads what the current scene is about even as tasks, objects, and layouts change — but it is slow to react to events as brief as a contact transition. The first pins down when something changed; the second attaches which decision, in which phase, that change belongs to. (Method and experiments in progress.)
What is a critical phase?
Precision manipulation hinges on a handful of decision-sensitive moments. Consider driving a bolt into a hole (peg-in-hole). Following the classical contact-state decomposition (Mason 1981; Debus et al.), the task splits into contact regimes, and in each one a different control variable decides the outcome:
- Contact 1 — approach. Move too fast and inertia overcomes the magnetic hold; the bolt drops off the driver. Speed is decisive.
- Contact 2 — surface touch. Press too hard and the bolt bounces away. Contact force is decisive.
- Contact 3 — rim alignment. The bolt must move along the hole's normal direction. Vertical alignment is decisive.
- Contact 4 — seating. How far to turn / how hard to press to seat the bolt. Torque is decisive.
In each regime a small action error decides the success or failure of everything that follows. These intervals have historically been carved out and labeled by hand. We automate that in two steps:
- Phase segmentation. The VLA's generalization splits a rollout into phases. Boundaries like approach, contact, alignment and seating fall out without hand-written rules per task.
- Aligning with the failure distribution. We lay the points where the VLA failed on top of those phases and count how the failures spread across them. When failures pile up in one phase, we mark that phase as a critical phase.
Why critical phases?
Improving a frozen VLA — a pretrained policy whose weights are left untouched, with only a small module trained on top — has so far taken one of three routes:
- Spread RL uniformly over the whole rollout (SimpleVLA-RL). Reward comes from episode success alone and every timestep is swept the same way. With no idea where the trouble is, the budget gets spread evenly across the entire rollout.
- Have an expert rebuild the data (DAgger). Roll the policy out, have an expert label the correct action on the states it actually visited, add those to the dataset, retrain. When the labeling stops, so does the improvement.
- Have a human designate the segment and run RL only there (RLT). A lightweight actor-critic head sits on a compact readout — the RL token — from the frozen VLA and trains online on a real robot. Across four real-robot tasks (screw installation, zip-tie fastening, charger insertion, Ethernet insertion) it speeds up the hardest segment by up to 3×.
RLT already validates the premise of this work: concentrate the budget on the hard segment and the policy really does improve fast. The catch is that a human picks the segment. RLT marks the insertion / fastening / rotation segment up front, a human operator chooses at which point in each episode to hand control from the base VLA to the RL policy, a human labels every episode success or failure at the end (rT = 1 or 0), and a human takes over by teleoperation when the robot gets stuck.
All three stall in the same place: the system cannot tell on its own when or where things go wrong. The first doesn't know, so it searches the whole rollout; the other two are told by a human. Try to close the loop so that the VLA improves itself, and an expert is still standing inside the loop.
This is the gap critical phase detection fills. Once the system can point out where the outcome is being decided by itself, the "hand over to RL here" moment that RLT asks a human for comes for free — and the human steps out of the loop.
Contributions
1. Definition & detection
Critical phase formalized as decision sensitivity, with a per-timestep score ct learned from success rollouts only. Validated against counterfactual ground-truth decisiveness — the measured rate at which perturbing the action at t flips the outcome in simulation.
2. Criticality ≠ failure
Direct evidence that ct is not a failure probability: it responds on success rollouts where failure detectors stay flat, and it catches the decision point of spatially-overlapping failures they miss. Head-to-head with FAIL-Detect, FIPER, SAFE on LIBERO.
3. Localized RL refinement
Concentrating reinforcement-learning refinement on the detected critical phases accelerates a frozen VLA's convergence on precision tasks, reaching higher sample efficiency than uniform RL fine-tuning (VLA-RL, SimpleVLA-RL).
BibTeX
@article{anonymous2026cpd,
title = {Self-Supervised Critical Phase Detection for VLA Refinement},
author = {Anonymous},
journal = {Preprint},
year = {2026},
note = {Under review},
}