We use aloha to make real data on real task like Grasping Cube into Bowl.
We collected about 200 episodes, each episode about 500 steps.
We use simple net config: jsut resnet18 for image, MLP for qpos, unet1d for dolicy, and DDPM noise scheduler from diffusers with default config from colab demo (https://colab.research.google.com/drive/18GIHeOQ5DyjMN8iIRZL2EKZ0745NLIpg?usp=sharing).
We trained 1000 epochs with all four views images. Sadly, it DO NOT WORK AT ALL.
PS: obs horizon 2, predtion action 16, action horizon 8
@cheng-chi @pointW hi guys, Would you please give some advices?
huangxu1991
commented
2 months ago
We use aloha to make real data on real task like Grasping Cube into Bowl. We collected about 200 episodes, each episode about 500 steps. We use simple net config: jsut resnet18 for image, MLP for qpos, unet1d for dolicy, and DDPM noise scheduler from diffusers with default config from colab demo (https://colab.research.google.com/drive/18GIHeOQ5DyjMN8iIRZL2EKZ0745NLIpg?usp=sharing). We trained 1000 epochs with all four views images. Sadly, it DO NOT WORK AT ALL. PS: obs horizon 2, predtion action 16, action horizon 8 @cheng-chi @pointW hi guys, Would you please give some advices?