Researchers from Qianxun Intelligent's Gaoyang Team have developed Legato, a novel training method for Vision-Language-Action (VLA) models that enables robots to perform actions with natural, smooth transitions. Unlike previous methods that patched continuity during inference, Legato integrates continuity directly into the training process. This is achieved through a noise-real value mixing mechanism and a step-by-step denoising process that ensures consistency between training and inference, leading to more fluid and efficient robotic movements. AI
IMPACT Enables more fluid and efficient robotic movements by addressing continuity issues in action generation, potentially accelerating adoption in complex manipulation tasks.
RANK_REASON The item describes a new training method for robots presented in a research paper accepted to a conference. [lever_c_demoted from research: ic=1 ai=1.0]
- Action chunking as conditional policy compression
- Flow Matching for Generative Modeling
- Gaoyang Team
- Legato
- Qianxun Intelligent
- Real-Time Chunking
- RSS 2026
- Vision-language-action model
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
