TeleIllusion: Foundational Robot Manipulation Simulation Platform

Published:

**TeleAI, Shanghai09.2025 - 06.2025**
  • Designed and implemented an asynchronous parallel simulation framework based on Isaac Lab, integrating skill-based motion generation and comprehensive domain randomization across lighting conditions, visual appearance, spatial configurations, and camera configurations.
  • Developed a distributed data generation pipeline and deployed it into a compute cluster for large-scale data synthesis, synthesizing over 144 hours of manipulation demonstrations per day.
  • Trained the Splat VAE and DiT within the Trellis framework on self-curated articulated object datasets to generate object part meshes, kinematic structures, and manipulation affordances simultaneously.