3 ms·
Stanford's HumanPlus Robots: Humanoid Shadowing and Imitation from Humans
- Jimmc414 2y agoA low-level policy or NN controller called Humanoid Shadowing Transformer is trained via reinforcement learning in simulation using an existing 40-hour human motion dataset (AMASS). This policy allows the humanoid to shadow/imitate human motions in real-time using just an RGB camera. Through the real-time shadowing, human operators can then teleoperate the humanoid to collect egocentric vision data of the humanoid performing various tasks. This collected real-world data is then used to train vision-based high-level policies called Humanoid Imitation Transformers via supervised behavior cloning. These policies enable the humanoid to autonomously perform the tasks. Number of human demonstrations were needed per task: Wear a Shoe and Walk: 40 demos Fold Clothes: 40 demos Rearrange Objects: 30 demos Type "AI": 30 demos Two-Robot Greeting: 30 demos "Warehouse": 25 demos According to bill of materials the robot can be 3D printed and assembled for $107,945.41 [0] [0] https://docs.google.com/document/d/1SOrTv9INAKD5Yd9oLjfND1N75sRkg6ILJWmE74H1zJ8/edit#heading=h.68rue25vhirl https://docs.google.com/document/d/1SOrTv9INAKD5Yd9oLjfND1N7...
- verdverm 2y agoThis is a training improvement over the Aloha project from the same group, which iirc, got to about 80% accuracy after 90 training examples