Summary

UniGoal enables universal zero-shot goal-oriented navigation, allowing agents to reach arbitrary goals in diverse environments without task-specific training. It targets embodied AI and robotics researchers/developers, offering flexible navigation via a unified graph representation.

How It Works

The core innovation is a unified graph representation for navigation goals and environments, enabling direct generalization to unseen scenes and goal types without retraining. The method integrates visual perception, language understanding, and spatial reasoning for goal decomposition and path planning.

Quick Start & Requirements

• Installation: Requires Python 3.8 and Conda. Clone repo, create unigoal Conda env, install dependencies (habitat-sim, habitat-lab, pytorch3d, detectron2, LightGlue, Grounded-Segment-Anything, GroundingDINO, faiss-gpu). Download SAM and GroundingDINO models.
• Datasets: HM3D scenes, instance-image-goal episodes, text-goal episodes are necessary.
• LLM/VLM: Supports local Ollama (ollama pull llama3.2-vision) or custom API configuration.
• Execution: Evaluate via python main.py --goal_type <ins-image|text>. Infer using goal type and details (image path/text). Real-world deployment needs implementation in src/envs/real_world_env.py.
• Links: Paper, Project Page, Video.

Highlighted Details

• Zero-Shot Generalization: Navigates to novel goals/environments without prior training.
• Multi-Modal Goals: Supports instance images and natural language descriptions.
• Research Advancements: Extends prior work with GC-VLN (CoRL 2025), using scene graphs for V&L navigation.
• Simulation & Real-World: Framework for Habitat simulation and outlines real-world integration steps.

Maintenance & Community

Active development is indicated by recent updates and conference acceptances (CVPR 2025, CoRL 2025). The README lacks community channel links, roadmaps, or detailed contributor information.

Licensing & Compatibility

The repository's license is not specified in the README. This omission is a significant adoption blocker, especially for commercial use or integration into proprietary systems.

Limitations & Caveats

Object-goal navigation support is pending. Real-world integration requires manual implementation of environment-specific functions. As a research project, stability and long-term maintenance are TBD. The unspecified license is the primary barrier to assessing usage rights.