At XAIR 2026, AutoLife previewed its Particle-Grounded Differentiable World Model and showed progress on the S3.
On September 4, 2026, at XAIR 2026 in Guangzhou, AutoLife previewed its Particle-Grounded Differentiable World Model, or PGD World Model, and shared validation progress on its S3 embodied-intelligence robot.
AutoLife positions the model as a foundation for general embodied intelligence. It connects real-scene reconstruction, object-level editing, physical reasoning and native differentiable optimization, turning a real environment into a scene that a robot can compute, edit and simulate.
Using short multi-view videos, the system reconstructs spatial structure and appearance, then represents objects with position, shape, material and physical state. Engineers can adjust pose, scale, material, mass, friction and hardness, or add and replace objects to create controlled, repeatable task conditions.
The work builds on Dr. Chen Siwei’s research and AutoLife’s engineering program. His first-author paper, “DaXBench: Benchmarking Deformable Object Manipulation with Differentiable Physics,” was selected as an ICLR 2023 Oral presentation and established a foundation for differentiable-physics manipulation research.
The team is now connecting scene reconstruction, a unified particle representation, physical processes and robot data loops around real training tasks. Founder, CEO and CTO Dr. Chen Siwei says robots need to understand not only language and images, but also how materials, contact and actions change the physical world.
S3 combines dexterous arms, laser-vision navigation and multimodal speech. With the PGD World Model, it can rehearse actions, update its scene model and replan motion for changing indoor settings such as supermarkets, dining, reception, home and retail.
At the expo, AutoLife demonstrated popcorn preparation and delivery, hands-on home scenarios, and smart retail tasks including replenishment, assisted shopping and checkout.
AutoLife will continue refining the world-model stack, using real commercial environments to build a data flywheel and move embodied intelligence toward repeatable, large-scale deployment with industry partners.