data

command and state

kuavo_traj_q_v_tua

policy

module

robot env

Chat with HW

experience

事务

Idea

1. 图像的归一化能否和state一样在图像之间归一化，而不是自己归一化？

• what matters in robot imitation learning ?

• diffuision scheduler and predict_net is 解耦

• directly use data or use latent of data

• some acknowledges about nn.parameters()

• a model archtecture picture about diffusion policy

• details about imitation learning data flow in training

• naive bayes 在对两个高斯分布时，naive bayes 是logistic回归的特殊情况

• 什么是自归回

• nn.Conv1d(in_channels, out_channels, kernel_size, stride, padding), nn.Conv2d, nn.Linear(),

• Conditional Denoise:

• VAE IL, DIFFUSION IL, BC_RNN IL , (octo IL)?

关于部署端延迟以及模型预测时间的覆盖

条件：

1. 机器人数据的处理后频率为: 10Hz
2. 模型预测时间约为: 0.13s

设计

1. 机器人轨迹设计为：1-100step, 时间为10s(10Hz)

延迟模拟