HorizonRobotics/HoloAgent

What it solves

HoloAgent provides a unified framework for general-purpose robots to perform complex, long-horizon tasks in the real world. It addresses the challenge of connecting high-level language instructions to physical execution while maintaining a spatial understanding of the environment through 3D memory.

How it works

The system operates as an "Embodied AgentOS" that converts language instructions into monitored skill graphs. It integrates four key components:

• Embodied AgentOS: Manages high-level task planning, runtime feedback, and closed-loop execution.
• 3D Spatial Memory: Provides physical-world spatial representations to ground robot reasoning for long-term tasks.
• Embodied Skills: A library of executable navigation and manipulation skills that translate agent decisions into action.
• FSR-VLN: A specialized navigation component that uses a hierarchical multi-modal scene graph and "fast-and-slow" reasoning to handle long-range spatial navigation efficiently.

Who it’s for

Robotics researchers and developers building embodied AI agents capable of mobile manipulation, active exploration, and interactive human-robot interaction in dynamic real-world environments.

Highlights

• Closed-Loop Execution: Integrates spatial retrieval, execution, memory updates, and recovery mechanisms to ensure tasks are completed successfully.
• 3D Spatial Reasoning: Uses 3D memory to allow robots to explore new spaces and update their understanding of the environment online.
• Multi-Robot Coordination: Supports coordinating navigation and motion across multiple robots, such as humanoid robots.
• Hierarchical Navigation: Implements FSR-VLN for efficient vision-language navigation using multi-modal scene graphs.