@Rocwng Agreed that recovery matters as much as zero-shot. For now, the blog shares per-task scores, including the tasks where Dream-1 falls short: https://t.co/RdpdhNDHr8
@vaibhavmehtain@robocurve Good questions. We haven't published a breakdown of where that time goes yet, but it's a fair ask and we're looking at sharing it. Today Dream-1 runs in the cloud and is built for tabletop manipulation. Faster control loops come into play as we expand to mobile manipulators and https://t.co/ySzsxA6vAX
@s_arkhangelskiy Dream-1 is focused on two-arm, parallel-gripper robots for now. Single-arm support isn't part of this release, but we're noting the interest.
@JackSk35800 Good question, and it's a real limitation. On unseen robots, Dream-1 predicts gripper poses and the robot's IK solver executes them, so errors tend to show up near obstacles, at joint limits, or on bodies far from what it has seen. When zero-shot falls short, a 5-minute demo from
@Yvan_Ciz Zero-shot on 27 unseen robots in sim, scores range from about 92% down to about 32% of what it gets on YAM, the robot it was trained on. We also report real-world unseen robot performance as well. A lot depends on the robot's IK solver. Full breakdown: https://t.co/RdpdhNEfgG
@thehypedotnews Thanks! Inference runs in the cloud. Your robot sends camera observations to the API and gets actions back, and its own IK solver turns those into joint commands. No GPU needed on the robot.
