HuggingFace released ProjectSim, a system that estimates physical properties from images, on October 16, 2026, saying it can accurately predict mass, friction, and hardness from rendered views of 3D assets. It is the company's first research update since the release of the PhysGS benchmark in 2024.
ProjectSim reported a mean MnRE of 0.761 for mass, measured against scale measurements from the YCB dataset. That compares with the 0.654 score of the best published method, PhysQuantAgent.
ProjectSim is built on a vision-language model and targets applications in robotics and physics simulations. Availability begins with research access, initially for academic and industry researchers.
"We evaluated ProjectSim’s image-based physical property estimates against measurements from three research datasets: YCB, NeRF2Physics and IDPP," said Luca Cilio, projectsim contributor. The comparisons cover mass, sliding friction, indentation hardness and restitution, which describes rebound at impact.
The announcement follows the release of the PhysGS benchmark, which HuggingFace said frames the significance of ProjectSim's results as a step toward more accurate physics simulations in virtual environments.
HuggingFace did not say how ProjectSim performs on dynamic interactions, and raised the open question of whether the model can generalize to unseen objects. The company said it will release further results on the benchmark in the coming months.
Source: huggingface
