ICI-VLA

“ICI-VLA: In-Context Imitation with Spatiotemporally Aligned Demonstrations for Vision-Language-Action Models”

“ICI-VLA updates its parameters only during offline training; at inference, the policy remains fixed and conditions action generation on retrieved micro-demonstrations”

“on LIBERO-Long, ICI-VLA reaches 96.8% vs. 93.2% for OpenVLA-OFT”

“π0 achieves 66.4% on physical tasks vs. ICI-VLA's 83.2%”

“Octo achieves only 12.5% average on RoboTwin 2.0 vs. ICI-VLA's 60.4%”