Research Index

Purpose

Collect research notes, benchmarking results, and references that inform ShadowHound system design.

Prerequisites

• Access to experiment logs and datasets.
• Understanding of current research priorities.

Steps

1. Create sub-pages for each experiment or literature review using the template.
2. Include reproducibility metadata such as dataset version, commit hash, and evaluation metrics.
3. Link promising findings back to roadmap items or implementation tickets.

Validation

• [ ] Each research note includes reproducibility metadata
• [ ] External papers are cited with accessible links
• [ ] Conversion pipeline preserves equations and figures

Architecture & Design Research

Foundation Models & Frameworks

• NVIDIA GR00T Analysis — Comprehensive analysis of Isaac GR00T foundation model for humanoid robotics, alignment with ShadowHound vision, and integration strategy

Key Findings: GR00T N1.5 (3B params) is a vision-language-action foundation model that aligns perfectly with our persistent intelligence approach. Uses Eagle 2.5 VLM + flow matching diffusion for cross-embodiment learning. Targets same hardware (Thor AGX). GR00T-Perception workflow explicitly includes "RAG memory" (validates our semantic memory approach!). Recommendation: Integrate GR00T as mission agent backbone while keeping DIMOS local planning + adding our spatial memory layer.

Persistent Intelligence (Multi-Brain Learning)

• Persistent Intelligence Architecture — Multi-brain architecture (Thor + Spark + Tower) with continuous learning cycles
• Early Design Priorities — Foundational patterns for Isaac Sim and future multi-brain deployment
• DIMOS Integration Analysis — Practical implementation mapping of persistent intelligence to DIMOS-Unitree framework

MVP Navigation & Perception (Oct 2025 Discovery)

• Persistent Intelligence MVP — Strategic roadmap building on original MVP with local planning first approach
• Local Planning Architecture — VFH + Pure Pursuit local planner architecture (navigation WITHOUT global maps)
• Hybrid Perception Architecture — YOLO + VLM integration patterns for embodied AI missions
• Local Planning Quickstart — Technical implementation guide for rapid MVP development

Key Insight: DIMOS local planner enables reactive navigation without SLAM, dramatically accelerating MVP (1 week vs 2-3 weeks). Local planning FIRST enables autonomous agent that can optionally use global maps when available. Reactive decisions create richer learning data for future persistent intelligence.

See Also

• Development Log — Research and development notes
• LLM Documentation — Local LLM research and benchmarks
• MVP Roadmap — Current MVP scope and milestones

References

• Documentation Root
• Research assets and datasets (link when available)
• Software Index
• Simulation Index