ACT: Action Chunking with Transformers on tinygrad

This repository contains an implementation of ACT (Action Chunking with Transformers) using tinygrad, a lightweight deep learning framework.

An overview of ACT with ALOHA (low cost bimanipulation hardware): https://tonyzhaozh.github.io/aloha/aloha.pdf

Features

• Implementation of ACT model architecture using tinygrad
• Support for simulated robotic manipulation tasks
• Training and evaluation scripts
• Integration with tinygrad's lazy evaluation and JIT compilation

How to use

Training

BEAM=2 DEBUG=2 python3.10 train.py

Eval

DEBUG=2 MUJOCO_GL=glfw python3.10 test.py