Event
HKUST proposes W2-VLA fine manipulation model
1 reports · 1 sources
Authors from the Hong Kong University of Science and Technology propose World-to-Wrist VLA (W2-VLA), which uses global task context to predict future wrist latent variables and then generates fine-grained actions based on them. The authors state that at inference time, neither future wrist observations nor autoregressive decoding of W2-CoT is required, and action generation can reach over 80Hz. The results reported by the authors are: LIBERO average success rate 98.5%, and RoboTwin 2.0's Easy/Hard settings at 60.71% and 18.21%, respectively; on three real-world tasks—tabletop cleanup, occluded placement, and dual-arm plug insertion—under both standard and OOD conditions, success rates and stage progress scores all exceeded all baselines.
Report timeline
- HKUST proposes W2-VLA: predicting future wrist latent variables enables 80Hz fine manipulation
Authors from the Hong Kong University of Science and Technology propose World-to-Wrist VLA (W2-VLA), which uses global task context to predict future wrist latent variables and then guide fine action generation. The method achieves 98.5% average success rate on LIBERO, 60.71% and 18.21% on the Easy/Hard settings of RoboTwin 2.0, respectively, and outperforms baselines under standard and OOD conditions on three real-world tasks: tabletop cleanup, occluded placement, and dual-arm plug insertion.
3D视觉工坊 · 微信公众号 ·