Code Update

Major bug fix by @charlesxu0124 in #38:

• Previously, inside the relative_env.py wrapper, the intervene_action executed by the robot (based on the spacemouse intervention action), was transformed in the wrong direction. It should have been transformed from the base frame to the end-effector frame, since the policy should learn actions in the end-effector frame and the spacemouse provide intervention actions in the robot base frame, but the transformation incorrectly applied was from the end-effector frame to the base frame. This caused experiments where the end-effector was not aligned with the robot base axis to not work.

The changes were made in the following files:
-serl_robot_infra/franka_env/envs/relative_env.py
-serl_robot_infra/franka_env/envs/wrappers.py

The updated code only populates the dictionary info with the "intervene_action" key-value pair if an intervention was provided. Otherwise this key does not exist. Other places in the code has been updated to check if "intervene_action" is in info before accessing it.

ResNet-10 weights pretrained on ImageNet-1k

This file contains weights for a ResNet-10 model pre-trained on ImageNet-1k data. It achieves 64.36% accuracy on validation data.

Franka Lift Cube Sim Demo 20 Trajectories (Image-based)

This file contains 20 successful lift cube trajectories, each with 100 steps. This can be used in the DrQ + RLPD sim example to bootstrap and accelerate learning.