1. FastTD3 Humanoid#

FastTD3 is a high-performance variant of the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm, optimized for complex humanoid control tasks. In this tutorial, we uses HumanoidBench, which have been integrated into RoboVerse. In this example we will set up and run FastTD3 within RoboVerse on the tasks of HumanoidBench.

Environment Setup#

# Step 1: Install FastTD3-specific requirements
cd RoboVerse/get_started/rl/fast_td3
pip install -r requirements.txt

# Step 2: Install RoboVerse with MJX simulator support
cd ../../..
pip install -e ".[mjx]"

One Command to Train FastTD3, Inference and Save Video#

We provide tutorials for training FastTD3, inference and saving video.

Run the following command to train a humanoid agent using FastTD3:

python get_started/rl/fast_td3/1_fttd3_humanoid.py

This script uses the following default configuration:

  • Simulator: mjx

  • Robot: h1

  • Task: humanoidbench:Stand

  • Environments: 1024 parallel instances

You can modify the task, robot model, or simulator by editing the CONFIG dictionary at the top of the script.

FastTD3 achieves fast and stable convergence: H1-Stand and H1-Walk tasks reach success threshold in under 10 minutes on a Quadro RTX 6000.

You can get the video like this:#

Stand:#

Walk:#

Run:#