Tag: Vision-Language-Action

Kyutai builds cheap memory into a 2.4B robot model

REKI keeps the last three frames inside its action expert using 32 register tokens, cutting arm oscillation from…