Embodied AI and robotics advances
Key Questions
What new models support embodied AI?
LingBot-Video uses MoE video pretraining to address domain mismatch for robotics. RoboDojo offers a unified sim-and-real benchmark across 42 simulation and 18 real tasks.
Where is real-world embodied AI being deployed?
Farming robots held an open house in Salinas featuring 15+ startups. This shows growing agricultural applications of embodied AI systems.
What research automates embodied agent design?
AgentCanvas and KDLoop explore automated architecture design for embodied agents. Limitations remain around rollout noise and local edit basins.
New open-source models and benchmarks for embodied AI: LingBot-Video (MoE video pretraining for embodied intelligence) tackles domain mismatch in video generation for robotics. RoboDojo provides a unified sim-and-real benchmark for generalist robot manipulation, covering 42 sim and 18 real tasks. Real-world deployment in agriculture: farming robots open house in Salinas with 15+ startups. New research on automating embodied agent architectures (AgentCanvas, KDLoop) shows promise but limitations like rollout noise and local edit basins. These developments signal growing investment in physical AI and robotics, with potential for defensible moats in evaluation infrastructure and domain-specific models.