rich annotations & update open-pi fsdp explanations

README.md

@@ -24,7 +24,7 @@
 InternDataEngine is a synthetic data generation engine for embodied AI that powers large-scale model training and iteration. Built on NVIDIA Isaac Sim, it unifies high-fidelity physical interaction from InternData-A1, semantic task and scene generation from InternData-M1, and high-throughput scheduling from the Nimbus framework to deliver realistic, task-aligned, and massively scalable robotic manipulation data.
 
 - **More realistic physical interaction**: Unified simulation of rigid, articulated, deformable, and fluid objects across single-arm, dual-arm, and humanoid robots, enabling long-horizon, skill-composed manipulation that better supports sim-to-real transfer.
-- **More task-aligned data generation**: LLM-driven task and instruction generation with task-oriented scene graphs (ToSG), producing structured scenes and rich multi-modal annotations (boxes, keypoints, trajectories) for complex instruction-following and spatial reasoning.
+- **More diverse data generation**: By leveraging the internal state of the simulation engine to extract high-quality ground truth, coupled with multi-dimensional domain randomization (e.g., layout, texture, structure, and lighting), the data distribution is significantly expanded. This approach produces precise and diverse operational data, while simultaneously exporting rich multimodal annotations such as bounding boxes, segmentation masks, and keypoints.
 - **More efficient large-scale production**: Nimbus-powered asynchronous pipelines that decouple planning, rendering, and storage, achieving 2–3× end-to-end throughput, cluster-level load balancing and fault tolerance for billion-scale data generation.
 
 ## 🔥 Latest News