markfzp/act-plus-plus

This repository contains code for training robots to imitate human movements using a technique called imitation learning. The primary target is Mobile ALOHA, a two-armed mobile robot system, but the code also includes two simulated table-top environments you can use to test things on a regular computer without any physical hardware. The repo implements three different learning approaches. ACT (Action Chunking with Transformers) is the main one, where the robot learns to predict a sequence of actions at each step rather than one action at a time, which helps it move more smoothly. Diffusion Policy is a second approach that frames the motion planning problem similarly to how image generators work, building up a motion plan by starting from noise and refining it. VINN is a third method. Each of these is a distinct way to turn a collection of recorded demonstrations into a policy the robot can follow on its own. The simulation environments use MuJoCo, a physics simulator, and include two tasks: picking up a cube and transferring it, and inserting a peg into a socket with both arms. You can generate training data by running scripted demonstrations in the simulator, train a model on those demonstrations, and then evaluate how well the trained policy performs on new runs. The success rate for the transfer cube task with ACT is described as around 90%, while the insertion task is around 50%. Installation requires Python 3.8 and a list of pip packages including PyTorch, MuJoCo, and dm_control. Training commands are run from the command line with flags to control batch size, number of training epochs, learning rate, and other parameters. A linked tuning guide notes that if the robot's movements come out jerky or it pauses unexpectedly, training for more epochs typically improves smoothness even after the training loss appears to have leveled off.