Research map / Vision-language and multimodal / Robot action and navigation
Cross-Hand Latent Representation for Vision-Language-Action Models
Guangqi Jiang, Yutong Liang, Jianglong Ye, Jia-Yang Huang, Changwei Jing, Rocky Duan, Pieter Abbeel, Xiaolong Wang, Xueyan Zou
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace raw joint-space actions with a shared latent action space in the VLA | pi0 · 10 real-world tasks, 4 dexterous hands | mean success rate: 0.32 -> 0.72 | Improved |
| use shared latent action space on the most distinct hand (XHand) | pi0 · XHand, 10 tasks | mean success rate: 0.29 -> 0.70 | Improved |
| zero-shot unseen task transfer with latent space versus pi0 plus kinematic retargeting | XL-VLA · unseen task-hand combinations | success rate | Improved |
| co-train xArm and G1 data in latent action space instead of raw actions | pi0 · G1 humanoid, 4 tasks | success rate: +57% | Improved |
| use unsupervised latent space instead of supervised LAD for latent replay | latent retargeting · Ability+Inspire replay | mean replay success: 0.60 -> 0.82 | Improved |
| use unsupervised latent space instead of LAD | latent retargeting · Paxini+XHand replay | mean replay success: 0.61 -> 0.81 | Improved |
| add retargeting loss L2 (distance and direction) (ablation) | latent autoencoder · cross-embodiment pinch direction error | direction error: 62.733 -> 11.857 | Improved |
| add retargeting loss L2 (ablation) | latent autoencoder · reconstruction joint RMSE | joint RMSE: 3.781 -> 5.476 | Mixed |
| add reconstruction loss L1 (ablation) | latent autoencoder · reconstruction joint RMSE | joint RMSE: 61.672 -> 5.476 | Improved |
| increase latent dimension from 32 to 128 (ablation) | latent autoencoder · latent space ablation | latent quality: 5.476 -> 5.324 | No clear effect |
| reduce latent dimension to 8 (ablation) | latent autoencoder · latent space ablation | reconstruction joint RMSE: 5.476 -> 20.913 | Hurt |
Closest papers by meaning
- Scaffolding Dexterous Manipulation with Vision-Language Models (NeurIPS 2025 · similarity 0.91)
- Cross-Embodiment Dexterous Grasping with Reinforcement Learning (ICLR 2025 · similarity 0.90)
- HandEdit: A Unified Benchmark for Egocentric Human-to-Robot Dexterous Hand Image Editing (NeurIPS 2026 · similarity 0.90)
- VLA-OS: Structuring and Dissecting Planning Representations and Paradigms in Vision-Language-Action Models (NeurIPS 2025 · similarity 0.89)
- RotVLA: Rotational Latent Action for Vision-Language-Action Model (NeurIPS 2026 · similarity 0.89)
- ManipTrans: Efficient Dexterous Bimanual Manipulation Transfer via Residual Learning (CVPR 2025 · similarity 0.89)
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.
