VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon
Yi Pan, Miao Pan, Qi Lu, Jiaming Huang, Man Zhang, Siteng Huang, Xin Li, Jie Zhang, Yongliang Shen, Xuhong Zhang, Wenqi Zhang
cs.RO
2026-07-02
ZJU and DAMO add a 40M latent monitor that truncates stale VLA chunks then gradient-guides replanning; π0.5 MetaWorld success 48.7%→64.4%, real-robot disturbance 40.0%→68.3%.
Generative VLAs emit a chunk of future actions in one forward pass, then execute a fixed action horizon open-loop so the expensive policy is called less often and motion stays smooth. A longer horizon buys fewer queries and a wider blind spot. Contact, slip, or a small pose error keeps executing stale actions until the drift is past recovery.
The paper measures that trade-off. Stretching the horizon on π0.5 cuts policy calls by about 4× and drops success from about 64% to below 49%; SmolVLA and X-VLA move the same way. H=1 restores reactivity and destroys the reason for chunking, because the VLA runs at every control step. A single static horizon is also hard to pick in advance: the sweet spot moves with task difficulty, dynamics, and sim-to-real gap.
VLA-Corrector is an inference-time plugin. Backbone weights stay frozen. After the VLA is trained, its visual encoder is frozen on benchmark demonstrations and a 40M MLP learns to predict the short-horizon visual latent residual from the current latent and the action just executed. The loss mixes L2 with cosine direction. The module is not a full world model. It only learns how the scene should evolve when execution is still on track. Teleop jitter in the demos is acceptable if the trial as a whole still advances the task.
At deploy time the Latent-space Vision Monitor scores mismatch as Et = 1 minus cosine similarity between predicted and observed residuals. A sliding-window median and MAD set adaptive on/off thresholds; an interrupt fires only after several consecutive steps above Ton, so a single noisy frame does not cut the chunk. Remaining queued actions are dropped, and the fixed horizon becomes an event-triggered adaptive horizon.
Truncation alone is not the recovery. Online Gradient Guidance runs on the single policy call after an interrupt. It subtracts accumulated drift from the pre-drift expected residual, then pulls the visual effect of the candidate action toward that corrective direction by injecting the gradient into the flow-matching velocity field. Default η is 1. The monitor has to stay outside the policy: an auxiliary residual head fine-tuned on the backbone reaches only 49.55% average success, while the external LVM reaches 64.35%. The internal head rewrites representations used for planning; the external head trains on frozen features and leaves the policy alone.
On MetaWorld, all three backbones improve, and harder splits improve more.
| backbone | baseline avg success | +Corrector |
| π0.5 | 48.70% | 64.35% (+15.65; Very Hard 41.0% to 65.0%) |
| SmolVLA | 61.90% | 66.65% (+4.75) |
| X-VLA | 55.55% | 59.60% (+4.05) |
The horizon sweep does not buy success by querying more. At H=50, π0.5 goes from 48.72% to 58.70% success with calls per episode 5.15 to 4.98, a 24.6% success-per-call gain. At H=10, SmolVLA goes from 61.90% to 73.00% with calls 19.27 to 15.64, a 45.3% efficiency gain. On LIBERO, few-shot π0.5 plus Corrector reaches 97.80% average, above the fully fine-tuned 96.95%. On a real AgileX PiPER 6-DoF arm, three groups of three tasks with 20 trials each rise from 55.6% to 73.3% average; disturbance recovery is the largest jump, 40.0% to 68.3%.
Ablations: truncation alone reaches 60.35%, adding OGG reaches 64.35%. 83.7% of truncations fall in critical phases such as grasp and alignment, 16.3% in tolerant phases. After an interrupt, OGG raises 10-step recovery by 0.23 on average. A corrector trained on 60% of demos already beats the baseline. A LIBERO-trained corrector transferred to MetaWorld only moves 48.7% to 51.8%; the in-domain corrector reaches 58.7%. Enabling OGG costs about 1.62–1.68× wall-clock, or +7.93 ms per environment step once amortized. A standard chunk query is about 278 ms; an OGG recovery query is about 588 ms.
Horizon is a dirty knob in VLA deployment: too short and compute dies, too long and contact tasks wobble. The result here is that a small monitor can decide whether the current chunk is still trustworthy, without editing the large model. For frozen π0.5, SmolVLA, and X-VLA this is a drop-in inference patch. Few-shot trajectories often cover the nominal path and miss drifted recoveries; extra fine-tuning data is not the only way to teach recovery.
This is an incremental patch, not a new VLA. It watches whether execution still looks like the demos. Failures outside that envelope stay failures.
The real-robot failure write-up is blunt. If the target leaves the workspace, the gripper is already in a bad pose, a tight insert has no force sensing, or vision is occluded or low-contrast, one corrective replan may not be enough. OGG can only bias actions the backbone already represents; it cannot invent a recovery skill the policy cannot express.
The monitor is trained on demonstrations, so it is biased toward what success looks like. Cross-domain transfer is only +3.1 points, and a new robot or camera should not be assumed to work. The 64.35% cross-architecture MetaWorld average is not the same aggregate as 58.70% at horizon 50 in the sweep; cite the setting. A 588 ms recovery query may be too slow for high-rate control. There is no head-to-head against methods that collect extra recovery data or train a separate recovery policy. Raising η to 10 or 100 hurts harder tasks, so strong guidance can yank the action off course.