CLoSD addresses multi-task character control by closing the loop between simulation and diffusion models, enabling complex character behaviors. It targets researchers and engineers in robotics, animation, and AI, offering a unified framework for tasks like motion imitation, sequence control, and text-to-motion generation.

How It Works

CLoSD integrates a Diffusion Planner (DiP), an autoregressive diffusion model, with a physics-based simulation environment (Isaac Gym). DiP generates motion trajectories that are then executed by a simulated humanoid character. This approach allows for real-time planning and control, leveraging the generative power of diffusion models for diverse motion tasks within a robust simulation loop.

Quick Start & Requirements

• Install: Create a Conda environment (conda create -n closd python=3.8, conda activate closd), install requirements (pip install -r requirement.txt), download and install Isaac Gym.
• Prerequisites: Python 3.8, Conda, NVIDIA GPU with ~4GB RAM (for running), ~50GB RAM (for training/evaluation), monitor (for running).
• Data/Models: Datasets (SMPL, AMASS, HumanML3D) and checkpoints are automatically cached. Adherence to their respective terms of use is required.
• Links: Project Page, Arxiv

Highlighted Details

• Reproduces ICLR 2025 Spotlight results for multi-task character control.
• Supports multi-task, sequence, and text-to-motion generation.
• Includes standalone DiP (Diffusion Planner) for motion generation and evaluation.
• Provides scripts for Blender visualization and SMPL parameter extraction.

Maintenance & Community

The project is associated with Guy Tevet and appears to be a research artifact from ICLR 2025. Further community or maintenance details are not explicitly provided in the README.

Licensing & Compatibility

• License: MIT LICENSE.
• Compatibility: Permissive for commercial use and integration with closed-source projects.

Limitations & Caveats

The setup requires specific versions of Python and Isaac Gym, and significant GPU RAM for training. The project is presented as a research artifact, and ongoing maintenance or support is not detailed.