UCAS Lab's ActionUNet Boosts VLA Robotic Success Rate by up to 11.4 Points in Hard Scenarios
机器之心 · wechat · 2026-11-23
Researchers from UCAS LAMP Lab propose ActionUNet, an efficient multi-scale fine-tuning framework that improves the robustness of pretrained Vision-Language-Action models after SFT without rebuilding the backbone.
Method
- Temporal U-Net models action features across coarse and fine time scales via down/up-sampling with skip connections;
- Conditional SIREN decodes actions as continuous time-queryable functions, weighted with Matérn kernels to reduce jitter;
- Efficient fine-tuning: the pretrained backbone is frozen; the module plugs into both flow-matching (π0.5) and regression-style (OpenVLA-OFT) VLAs, trains on a single RTX 4090, and adds only 6.46% inference latency.
Results
- RoboTwin2.0 Hard setting: π0.5 success rate rises from 29.8% to 39.6% (+9.8 pts);
- LIBERO-Plus: 65.8% → 71.9%, with 97.7% on standard LIBERO;
- Real dual-arm robot with distractors: 57.3% → 68.7% (+11.4 pts);
- Ablations show gains stem from multi-scale modeling, not smoothing or extra parameters.
Paper: arXiv 2609.34982; code open-sourced on GitHub.
More from Embodied
- ME has AI design a robot from photos of his motors and gears: "already a better engineer than me" — burhop · 2026-11-23
- Tesla Optimus shows off precise laminate flooring installation, piece by piece — ChuckDBrooks · 2026-11-23
- Two prompt lines about camera geometry took robot gripper control from 50% to 100% — YuXiang_IRVL · 2026-11-23
- Dyson R1 Nurovi arrives as a rare new robot vacuum in a market frozen since 2022 — Technovangelist · 2026-11-23
- Rutgers' ChatBCI pairs EEG with a local LLM, doubling ALS patients' typing speed — emmanuelvivier · 2026-11-23
- Siemens and AGIBOT generate robot grippers from text via agentic AI workflow — emmanuelvivier · 2026-11-23