PKU and Collaborators Introduce LDA-1B, a 1B-Parameter Embodied Foundation Model
chris_j_paxton · x · 2026-08-13
Peking University, alongside Galbot, CASIA, and others, has introduced LDA-1B, a 1-billion-parameter embodied foundation model set to be presented at RSS 2026.
- Data Scale: Trained on EI-30k, a dataset comprising over 30,000 hours of heterogeneous human and robot interaction data.
- Architectural Innovation: Jointly denoises action chunks and future DINO sequences within a unified multi-modal diffusion transformer framework, integrating visual forecasting, dynamics learning, and policy learning.
- Dynamics Modeling: Operates forward and inverse dynamics in a structured DINO latent space, avoiding redundant pixel-space appearance modeling to focus on task-relevant dynamics features.
More from Embodied
- Huawei's RoboHarness Orchestrates Heterogeneous Robot Policies Without Retraining — jiqizhixin · 2026-08-13
- JD.com Deploys Two-Wheeled Humanoid Robots in Supermarkets to Collect Teleop Data — CyberRobooo · 2026-08-13
- Stream's Ring-Maker Sandbar Says the Future of AI Wearables is Voice — TechCrunch AI · 2026-08-13
- Robotics Startups Are Everywhere, But Where Are the Actual Robots? — DJiafei · 2026-08-13
- Robotic Mower Keeps Unirrigated Grass Green Despite Heat and Drought — dreamwieber · 2026-08-13
- Community Deploys Hermes Agent on Raspberry Pi 4 with 2GB RAM for Low-Cost Always-On Agent — Teknium · 2026-08-13