Rohan Paul (@rohanpaul_ai)T2·78 pts
- V0.5 builds on the Wan2.2 video generator to jointly predict future camera frames and the motions producing them
- Under randomized clutter and lighting on RoboTwin 2.0 it drops only 7.6 points versus 24.7 for π0.5
- New industrial tasks are learned from under two hours of demonstrations, with weights fully open