TLDR: TANGO is a new robot navigation system that uses only an RGB camera to guide robots to object-level goals over long distances. It combines a global plan based on object connections with a local plan that considers traversable areas and avoids obstacles, all without needing complex 3D maps or prior training. It leverages AI foundation models for perception and includes a fallback mechanism for challenging situations, demonstrating improved performance and the ability to reach previously “seen but unvisited” locations.
Robots navigating complex, real-world environments has long been a significant challenge. Traditional methods often rely on creating detailed 3D maps, which can be computationally demanding and struggle to adapt to new, diverse settings. Alternatively, systems that learn to navigate might not fully use existing knowledge, limiting their efficiency.
Inspired by how humans navigate—by recognizing objects and places without needing precise 3D blueprints—a new approach called visual topological navigation has emerged. While image-level topological maps are simple, they often lack detailed semantic information and can be sensitive to changes in viewpoint. Object-level topological maps, however, offer advantages like natural language querying, semantic understanding, and consistent visual recognition, which are crucial for robots operating in open-ended environments.
Introducing TANGO: A Novel Navigation System
A new research paper introduces TANGO: Traversability-Aware Navigation with Local Metric Control for Topological Goals. This innovative system provides an RGB-only, object-level ‘topometric’ navigation pipeline that allows robots to navigate over long distances without prior training or the need for 3D maps. TANGO uniquely combines global topological path planning with local metric trajectory control, enabling robots to move towards object-level sub-goals while effectively avoiding obstacles.
The core idea behind TANGO is to continuously predict local trajectories using monocular depth and traversability estimation. It also includes an intelligent auto-switching mechanism that can fall back to a simpler controller when necessary. The system is built upon powerful ‘foundational models’ in AI, ensuring it can be applied to a wide range of scenarios without needing specific fine-tuning for each environment.
How TANGO Works
TANGO operates by bridging two distinct planning levels:
-
Global Topological Planning: The robot first builds a topological map of its environment, not with precise 3D coordinates, but by identifying and connecting meaningful objects (image segments). Using foundational models like Segment Anything Model (SAM) for segmentation and LightGlue for matching, it tracks objects across different views. When given a final object goal, TANGO uses a global planner (similar to a shortest path algorithm like Dijkstra’s) to determine a sequence of object sub-goals that lead to the final destination. This gives the robot a high-level understanding of ‘where to head’.
-
Local Metric Control: Once a topological sub-goal is identified, TANGO translates this into a local metric motion plan. It takes the robot’s current RGB image and uses Depth-Anything to estimate monocular depth, projecting the scene into a 3D Bird’s Eye View (BEV). Simultaneously, it uses CLIP (a vision-language model) to identify ‘traversable’ surfaces like ‘floor’ or ‘ground’ within the image. By combining this depth and traversability information, TANGO creates a local cost map. Dijkstra’s algorithm is then applied again to find the shortest, safest path through this cost map to the local sub-goal, generating precise control signals for the robot’s movement and steering.
A crucial feature is the Auto Switch Control. If the local metric traversability prediction becomes unreliable—for instance, if the robot is too close to a wall or an obstacle obstructs its view—TANGO automatically switches to a simpler, purely topological ‘segment servoing’ approach (based on RoboHop). This fallback mechanism ensures the robot can continue to make progress by rotating towards the goal until traversable regions are visible again.
Key Advantages and Capabilities
TANGO’s unique integration of global topological planning and local metric control offers several benefits:
-
Zero-Shot Navigation: It can navigate in new environments without requiring pre-trained controllers or extensive 3D maps.
-
RGB-Only: The system relies solely on a single RGB camera, making it resource-efficient.
-
Open-Set Applicability: By leveraging powerful foundation models, TANGO can understand and interact with a wide variety of objects and environments.
-
Obstacle Avoidance: The local metric control explicitly accounts for traversability, allowing the robot to navigate around obstacles effectively.
-
Reaching ‘Seen-But-Unvisited’ Goals: Beyond simple ‘teach-and-repeat’ tasks, TANGO can guide a robot to object goals it has observed during its mapping run but never explicitly visited. This demonstrates a significant step towards more generalized navigation.
Also Read:
- Enhancing AI Agents for Continuous Task Execution in Dynamic Environments
- DepthVision: Enabling Robots to See Clearly in Challenging Conditions with LiDAR-Enhanced Vision
Performance and Future Outlook
Experiments in simulated environments and real-world tests show that TANGO significantly outperforms existing state-of-the-art methods, including learning-based controllers like PixNav and other zero-shot controllers like RoboHop, across various trajectory lengths. The auto-switching control mechanism proved particularly effective, improving success rates in challenging scenarios.
While TANGO represents a major advancement, the researchers acknowledge limitations, particularly concerning perception errors (incorrect object matching) and the inherent geometric ambiguity of purely topological planning. However, the modular design of TANGO means that as more advanced perception models become available, they can be easily integrated to further enhance the system’s performance.
For more technical details, you can read the full research paper here.
