OpenSeeker and DOMINO: Pioneering AI Search and Robotics

🔍 AI Search

OpenSeeker's Open-Source notable advance

The Rundown

OpenSeeker just launched, marking a significant advancement in AI search capabilities. This fully open-source search agent leverages just 11.7k synthesized samples to achieve current best performance across multiple benchmarks. It outperforms existing models like DeepDive (29.5% vs. 15.3% on BrowseComp) and even industrial competitors like Tongyi DeepResearch (48.4% vs. 46.7% on BrowseComp-ZH). The model's two core innovations, fact-grounded scalable controllable QA synthesis and denoised trajectory synthesis, enable complex multi-hop reasoning tasks while maintaining high-quality action generation. By releasing the complete training dataset and model weights, OpenSeeker aims to democratize access to advanced search technology, fostering a more collaborative research ecosystem.

The details

• OpenSeeker was trained on just 11.7k synthesized samples, significantly less than conventional models.
• It achieved 29.5% accuracy on BrowseComp, surpassing DeepDive's 15.3%.
• On BrowseComp-ZH, OpenSeeker scored 48.4%, outperforming Tongyi DeepResearch's 46.7%.
• The model employs scalable controllable QA synthesis to enhance multi-hop reasoning tasks.
• Denoised trajectory synthesis promotes high-quality action generation from teacher LLMs.

Why it matters

OpenSeeker's release reflects a growing trend towards open-source AI, providing researchers and startups access to modern technology. This democratization can accelerate innovation and collaboration in AI search capabilities.

🤖 Robotics

DOMINO: Advancing Robotic Manipulation

The Rundown

DOMINO introduces a comprehensive dataset for robotic manipulation in dynamic environments. This large-scale benchmark encompasses 35 tasks with over 110K expert trajectories, addressing the limitations of existing vision-language-action models, which often struggle with moving targets. DOMINO's architecture, PUMA, integrates historical optical flow and world queries to forecast object-centric future states. In tests, PUMA achieved a 6.3% absolute improvement in success rates over baseline models. This advancement not only enhances dynamic manipulation but also ensures robust performance when transferring learned skills to static tasks, showcasing the potential for broader applications in robotics.

The details

• DOMINO features over 110K expert trajectories across 35 dynamic manipulation tasks.
• PUMA achieved a 6.3% absolute improvement in success rates over existing models.
• The dataset addresses the scarcity of dynamic manipulation datasets in the field.
• PUMA's architecture incorporates historical optical flow for enhanced prediction.
• Training on dynamic data improves spatiotemporal representations for static tasks.

Why it matters

DOMINO's introduction represents a significant leap in robotic manipulation, offering a pathway for more adaptable and capable robots. This could lead to practical applications in industries requiring real-time decision-making and interaction.