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

RotVLA: Rotational Latent Action for Vision-Language-Action Model

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 RotVLA, a VLA whose latent actions are continuous SO(n) rotation matrices learned with a triplet (three-frame compositional) objective, then used as a latent planner in a flow-matching action expert, reaching 98.2% on LIBERO and 89.6%/88.5% on RoboTwin2.0 with 1.7B parameters.

Problem

Existing latent action models (LAMs) use discrete quantization and encode-decode reconstruction, which leads to trivial frame-copying solutions, limited capacity, and latent spaces without physical meaning, hurting VLA pretraining on heterogeneous robot and human video data.

Approach

Model latent actions as SO(n) rotation matrices obtained by SVD projection of soft-quantized (SoftVQ) features, train with a triplet objective that enforces composition of two consecutive latent actions to predict the third frame, and jointly denoise latent and robot actions in a flow-matching expert so latent actions act as a high-level planner.

Key results

• RotVLA achieves 98.2% average success on LIBERO with 1.7B parameters, versus 98.1% for X-VLA and 96.6% for StarVLA.
• RotVLA reaches 89.6% (clean) and 88.5% (randomized) on RoboTwin2.0, above pi0.5 at 82.7% and 76.8%.
• On the LARY latent action benchmark, RotVLA achieves the lowest average regression MSE (0.20) and highest classification accuracy (70.98%), versus LAPA-DINOv3 at 0.63 and 45.62%.

Surprising findings

• The reconstruction-only LAM baseline yields latent actions that mostly encode appearance, as its imagined frame barely differs from the input frame.
• Adopting the SO(n) latent action alone gives only marginal gains; the compositional triplet supervision is what yields the additional improvement.

Limitations

• Model scale is moderate (1.7B parameters); the LAM has only 290M parameters, and scaling is left to future work.
• Human videos are used only as supplementary data, so the benefit of human-video-only pretraining is untested.
• Real-world evaluation is limited to three tasks on a single dual-arm platform with 30 trials per task.

Experimental setup

Datasets: Open X-Embodiment AGIBOT-beta RoboMIND RoboCOIN Ego4D LIBERO RoboTwin2.0 LARY benchmark

Benchmarks: LIBERO RoboTwin2.0 LARY real-world ARX R5 tasks

Metrics: success rate regression MSE classification accuracy inference latency

Baselines: GO-1 CoMo OpenVLA OpenVLA-OFT pi0 pi0.5 GR00T-N1.6 UniVLA villa-X X-VLA StarVLA Motus

Builds on: VQ-VAE SoftVQ UniVLA flow matching InternVL3.5 X-VLA

Closest papers by meaning

• Latent Action Pretraining from Videos (ICLR 2025 · similarity 0.92)
• Beyond World-Frame Action Heads: Motion-Centric Action Frames for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
• ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
• 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)
• Beyond Data Scaling: Representation-Centric Pre-training for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)