Research map / Vision-language and multimodal / Robot action and navigation
ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning
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 ActionUNet, a lightweight multi-scale fine-tuning module (temporal U-Net in action feature space plus a conditional SIREN decoder with Matérn-weighted smooth aggregation) that improves frozen pi0.5 and OpenVLA-OFT policies on RoboTwin 2.0, LIBERO, LIBERO-Plus, and real-world dual-arm tasks.
Problem
VLA models map coarse multi-modal semantics directly to fine-grained action sequences, causing a temporal scale mismatch that hurts robustness to distractors and perturbations in cluttered environments.
Approach
Inject multi-scale temporal priors via a lightweight temporal U-Net over the action feature space, then decode with a conditional SIREN whose time-dependent biases and Matérn-5/2 weighted aggregation enforce smooth, C2-continuous action trajectories, while the VLA backbone stays frozen.
Key results
- On RoboTwin 2.0 (12 tasks), pi0.5+ActionUNet improves average success from 59.6% to 68.4% (easy) and 29.8% to 39.6% (hard).
- On LIBERO-Plus, pi0.5+ActionUNet raises average success from 65.8% to 71.9%; on LIBERO, pi0.5 goes from 95.6% to 97.7%.
- Real-world average success rises from 76.0% to 88.7% (easy) and 57.3% to 68.7% (hard); policy call latency increases by about 6.5% (60.9 to 64.9 ms).
Surprising findings
- The U-Net alone slightly lowers easy-setting success (69.0% to 67.5%) while the SIREN alone lowers hard-setting success (34.5% to 33.3%), so the two components are complementary.
- The Matérn kernel adds a small further gain only once combined with U-Net and SIREN; U-Net plus SIREN without it already gives most of the hard-setting gain.
Limitations
- Uses fixed temporal scales and models action chunks independently, per the authors.
- Scale-aware modeling is applied only at the downstream action stage, not during VLA pre-training.
- Real-world evaluation covers only three tasks on one platform.
Experimental setup
Datasets: RoboTwin 2.0 LIBERO LIBERO-Plus
Benchmarks: RoboTwin 2.0 LIBERO LIBERO-Plus Realman real-world dual-arm
Metrics: success rate policy inference latency
Baselines: pi0.5 OpenVLA-OFT pi0 RDT ACT DP3 WorldVLA SmolVLA UniVLA VLA-Adapter RTC
Builds on: SIREN flow matching U-Net Matérn kernel
Closest papers by meaning
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.93)
- ATI-VLA: Action-Centric Predictive Vision–Language–Action Models via Actionable Alignment Then Adaptive Injection (NeurIPS 2026 · similarity 0.92)
- VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model (ICML 2026 · similarity 0.92)
- VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching (NeurIPS 2025 · similarity 0.92)
- ForceVLA: Enhancing VLA Models with a Force-aware MoE for Contact-rich Manipulation (NeurIPS 2025 · similarity 0.91)
- VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.