@ghaffarqureshi40: NVIDIA is training robots using Apple Vision Pro?! NVIDIA's Project GROOT is revolutionizing how we train humanoid robots, they've developed a mind-blowing pipeline that multiplies data by 1000x or more using simulation. 1. Teleoperation with Apple Vision Pro: Using the Apple Vision Pro, human operators can control humanoid robots in real-time, providing immersive first-person control. This allows to collect demonstration data, even though it's time-consuming. 2. RoboCasa Simulation Framework: RoboCasa multiplies this demonstration data by creating diverse virtual environments. For example, a single kitchen setup can be transformed into hundreds of unique ones in simulation, varying textures, furniture, and object placements. 3. MimicGen Augmentation: MimicGen takes it further by generating new robot motion trajectories based on the initial human data. It filters out unsuccessful attempts, creating a vast dataset from limited human input. Traditionally, training robots relied on slow, expensive human data collection. Now, we can trade compute power for that time, allowing robots to learn much faster and in a wider range of situations. This approach addresses the limitations of teleoperation, which is bound by the physical constraints of time and space. By shifting to a simulation-based model, we're trading expensive human data for compute power, breaking barriers in robot training and adaptability. This method could lead to rapid advances in robot capabilities, akin to the breakthroughs we've seen with large language models. #creatorsearchinsights

1363

There are no more comments for this video.

To see more videos from user @ghaffarqureshi40, please go to the Tikwm homepage.