Research map / Vision-language and multimodal / Robot action and navigation

VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon

NeurIPS 2026 · Vision-language and multimodal › Robot action and navigation · approach: Vision-language-action foundation models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes VLA-Corrector, a frozen-backbone inference wrapper that monitors latent visual dynamics with an external ~40M MLP, truncates stale action chunks on persistent drift, and steers the next flow-matching replan with online gradient guidance.

Problem

Fixed action horizons in action-chunked generative VLA policies trade policy-call efficiency against open-loop blindness, so stale actions accumulate errors before the next replan.

Approach

An external latent-dynamics corrector predicts the expected visual residual from the executed action; a cosine mismatch score with robust MAD-based hysteresis triggers interrupt-based truncation, and the next replan is guided by a gradient that aligns predicted action effect with a corrective latent direction.

Key results

• On MetaWorld with pi0.5, average success rises from 48.70% to 64.35% (Easy-to-Very Hard splits), and on LIBERO a few-shot model with VLA-Corrector reaches 97.80% vs 96.95% for full fine-tuning.
• On pi0.5 at horizon 50, success rises from 48.72% to 58.70% while policy calls drop from 5.15 to 4.98 (+24.6% success-per-call).
• Real-world AgileX PiPER average success improves from 55.6% to 73.3%, with the largest gain (+28.3) on disturbance recovery.

Surprising findings

• A coupled detector (auxiliary head fine-tuned into the VLA) performs worse than the decoupled external LVM (49.55% vs 64.35%), suggesting backbone-coupled monitoring harms action generation.
• Enabling OGG increases wall-clock inference time by about 1.64x, though it is event-triggered.

Limitations

• OGG adds about 2.1x latency per recovery query and 1.64x overall inference time.
• Recovery is bounded by the frozen backbone's action prior; cannot create behaviors the policy cannot represent.
• Real-world evaluation is small (9 tasks, 20 trials each) and failures occur when disturbances move targets out of reach.

Experimental setup

Datasets: MetaWorld demonstrations LIBERO demonstrations AgileX PiPER real-world demonstrations

Benchmarks: MetaWorld LIBERO AgileX PiPER real-world tasks

Metrics: success rate policy calls per episode success-per-call efficiency post-interrupt recovery rate inference time

Baselines: pi0.5 open-loop baseline SmolVLA X-VLA coupled internal detector + OGG

Builds on: flow matching classifier guidance action chunking

Closest papers by meaning

• ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.93)
• VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching (NeurIPS 2025 · similarity 0.93)
• EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.91)
• ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning (NeurIPS 2026 · similarity 0.91)
• ATI-VLA: Action-Centric Predictive Vision–Language–Action Models via Actionable Alignment Then Adaptive Injection (NeurIPS 2026 · similarity 0.91)
• VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model (ICML 2026 · similarity 0.91)