原帖引用 · 原文
Rohan Paul@rohanpaul_ai
V0.5 trains on two objectives at once: predicting the robot’s next actions and predicting what its cameras will see next.
The video prediction is only there to teach the model how scenes change.
At runtime, V0.5 outputs actions directly, so the policy gets a world model’s scene understanding without generating video on every control step.
Brilliant release by @VerneRobotics
V0.5, an open-weight world action model, that looks through cameras and outputs the next second of arm, gripper and mobile-base motion.
The big deal is that it learned new industrial tasks from under 2 hours of demonstrations.
> Robot models usually stumble when a room changes. With clutter and new lighting that neither model trained on, Verne Robotics’ V0.5 lost 7.6 points in simulation, while Physical Intelligence’s π0.5 lost 24.7.
> On RoboTwin 2.0, randomized clutter, lighting and textures cut V0.5 from 81.0% to 73.4% in Verne’s tests, while Physical Intelligence’s π0.5 fell from 70.7% to 46.0%.
> Built on the Wan2.2 video generator, V0.5 learned from over 1.7M demonstrations to predict upcoming camera frames alongside the motion producing them.
https://t.co/BaTwYMcNsj
https://t.co/8WKzZHXJIw
在 X 查看这条帖子 ↗
Introducing V0.5, Verne’s open-weight world action model for industrial robots.
We believe democratizing robotics starts with making frontier foundation models open to everyone.
Bringing robots to every industry will take many teams, building on shared progress across many workflows.
V0.5 is the leading US-made open-weight model on LIBERO and RoboTwin 2.0 benchmarks.
We hope you’ll join us in putting V0.5 to work.
在 X 查看这条帖子 ↗
· UTC+8