Research map / Vision-language and multimodal / Robot action and navigation
Beyond World-Frame Action Heads: Motion-Centric Action Frames for Vision-Language-Action Models
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 MCF-Proto, a lightweight VLA action head that predicts an observation-conditioned rotation defining a local motion-centric frame and composes actions from shared prototypes, improving LIBERO and LIBERO-plus robustness.
Problem
VLA action heads typically regress end-effector actions in a fixed world frame, which may be a poor match for local manipulation structure and brittle under geometric perturbations.
Approach
Predict a per-step rotation R_t in SO(3) defining a Motion-Centric Action Frame, compose local translation and rotation actions as soft mixtures of shared learned prototype matrices, and map them back to the world frame, trained end-to-end with only standard action supervision plus orthogonality and temporal-smoothness regularizers.</price>
Key results
- On LIBERO, MCF-Proto averages 97.7% success vs. 97.1% for OpenVLA-OFT (best baseline).
- On LIBERO-plus, Camera 69.7% and Robot 66.0% vs. OpenVLA-OFT 59.7% and 37.2%.
- Ablation: MCF + Proto reaches 96.9% avg vs. 93.8% for BC MLP head; world-frame Proto alone reaches 94.4%.
Surprising findings
- Prototype composition in the world frame alone gives less gain than in the MCF, and the paper reports the frame and prototypes as mutually reinforcing.
Limitations
- Evaluation limited to LIBERO, LIBERO-plus, and a small number of real-world tasks.
- Learned frame is an implicit latent representation, not semantically grounded.
- Depends on coverage and quality of offline demonstrations.
Experimental setup
Datasets: LIBERO LIBERO-plus
Benchmarks: LIBERO LIBERO-plus OpenArm real-world tasks
Metrics: success rate average length angular error covariance trace effective rank PCA top-3 explained variance
Baselines: pi0+FAST OpenVLA-OFT pi0 FLOWER GR00T-N1.5 BEAST OpenVLA pi0-fast Nora WorldVLA UniVLA RIPT-VLA
Builds on: Qwen3-VL OpenVLA-OFT
Closest papers by meaning
- RotVLA: Rotational Latent Action for Vision-Language-Action Model (NeurIPS 2026 · similarity 0.92)
- ATI-VLA: Action-Centric Predictive Vision–Language–Action Models via Actionable Alignment Then Adaptive Injection (NeurIPS 2026 · similarity 0.92)
- ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning (NeurIPS 2026 · similarity 0.91)
- NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · 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.