Rhoda AI demos FutureVision shell game

Interesting robotics research, but the real test is whether the gains hold beyond curated demos and narrow tasks.

• –The web-video pretraining angle is compelling because it tries to borrow scale from internet video instead of relying only on expensive robot data.
• –The shell game is a sensible stress test for object permanence and memory, since the robot has to track state through occlusion and repeated shuffles.
• –The claim of strong task learning from roughly 10-20 hours of robot data is the most commercially relevant part; if reproducible, it could lower the cost of adapting policies to new embodiments.
• –The system still depends on an inverse-dynamics translator, so the “video-to-action” story is not the same as a single end-to-end policy in production.
• –This reads more like a serious research milestone than a product launch, and it will need independent validation, broader benchmarks, and more failure analysis before anyone should treat it as robust.