日拱一卒王小楼
2025 Robot research explodes
Architectural and training challenges are overcome
Data collection, data quality, recipe validation
With RL support, Figure/Dyna/PI success rate >99% in the field
Self-improvement and self-recovery frameworks mature
VLA fine-tuning approaches, retaining generalists → integrating expertise into generalists
Action segmentation, FAST tagging
Robot actions no longer lag, approaching human speed
Multimodal fusion: vision/speech/touch
Tactile feedback compensates for visual deficits, significant improvement
View OriginalArchitectural and training challenges are overcome
Data collection, data quality, recipe validation
With RL support, Figure/Dyna/PI success rate >99% in the field
Self-improvement and self-recovery frameworks mature
VLA fine-tuning approaches, retaining generalists → integrating expertise into generalists
Action segmentation, FAST tagging
Robot actions no longer lag, approaching human speed
Multimodal fusion: vision/speech/touch
Tactile feedback compensates for visual deficits, significant improvement
:
