ResearchLow importance

VFlash generates visual plans for robots in one go and runs about 25 times faster

The most upvoted paper of the day on HF Papers (11/07). Instead of generating coordinates token by token, VFlash uses a lightweight “drafter” that predicts a whole point, box or trajectory in a single pass. On the 2D and 3D trajectory tasks tested, it reduces error compared with autoregressive decoding; preliminary measurements indicate about 25 times more speed in 3D motion forecasting. An action model guided by VFlash plans reaches 70.2% average success on four MolmoSpaces suites, versus 62.6% for the baseline.