TL;DR
This paper introduces SHOPPER, a robot navigation system that integrates vision, planning, and navigation as a service to task execution, rather than as a standalone goal. By assuming a structured but unknown environment, SHOPPER leverages semantic richness and passive mapping to efficiently navigate and complete tasks, demonstrating improved performance with experience.
What Problem Does This Paper Address?
Traditional robot navigation research often focuses on building accurate maps of environments as the primary objective. However, this approach becomes impractical for robots performing recurring real-world tasks. The authors argue that navigation should be a means to an end, supporting higher-level task execution. They identify a gap in systems that can efficiently navigate and perform tasks in structured, man-made environments where complete prior mapping is infeasible, but where general knowledge of the environment's organization (e.g., how items are arranged in a store) can be exploited.
Methods
SHOPPER's system is built around four core subsystems:
- Retriever: Given a goal item, this subsystem uses similarity metrics and the current map to determine a specific destination and a search plan for execution upon arrival. It leverages domain knowledge like item "type" (e.g., laundry detergents) and "counterpart" (e.g., pancake mix and maple syrup) relationships.
- Path Planner: This subsystem takes the destination from the Retriever and the robot's current location to plan a shortest path using Dijkstra's algorithm on a topological map. It annotates each node in the path with accumulated sensor information from past visits.
- Follower: This subsystem is responsible for executing the planned path. It uses vision algorithms to ensure the robot's current perception matches the predicted sensor information stored in the map, verifying its location and progress.
- Passive Mapper: This subsystem continuously updates a topological map of the world by monitoring the robot's actions and sensor readings during task execution. It focuses on distinctive places (intersections and places of interest) and their associated perceptual information.
A key innovation is the integration of a passive mapper with a planning system, where the map is built as a resource for the planner, rather than being an independent mapping program. The system also utilizes conditional sensing routines, meaning not all sensors are run continuously, which differs from previous approaches that relied on uniform, fixed-cost sensing.
Datasets
The paper uses a simulated environment called GROCERYWORLD. This simulator provides real color images from a local grocery store, with camera views restricted to four cardinal directions. The domain consists of 75,000 images. No external datasets are mentioned for training or evaluation.
Key Insights and Ideas
- Navigation as a Service: The most significant insight is reframing navigation not as a task in itself, but as a supporting service for higher-level task execution. This shifts the focus from creating perfect maps to enabling efficient task completion.
- Exploiting Environmental Structure: The system thrives on the assumption of a structured but unknown environment. This allows it to leverage general knowledge about how such environments are organized (e.g., item clustering in stores) to guide navigation and search.
- Passive Mapping for Efficiency: The passive mapper, which updates the map as a byproduct of task execution, is crucial. This avoids the overhead of dedicated exploration phases for mapping, making the system more efficient for recurring tasks.
- Semantic Richness in Perception: The system emphasizes semantically rich perception, using cues like signs and item types, rather than relying solely on low-level sensors like sonar. This richer perception is vital for both task performance and map construction.
- Experience Improves Performance: The paper demonstrates that as the agent performs more tasks, its learned map becomes more useful, leading to improved navigation efficiency.
Key Results
The paper demonstrates SHOPPER's ability to successfully find items in the GROCERYWORLD simulator. Specific quantitative results are not detailed, but qualitative findings include:
- SHOPPER successfully navigated and found items like "Solo laundry detergent," "Corn Pops cereal," and "Downy fabric softener" using different strategies (DEFAULT, SIGN-TYPE, COUNTERPART).
- The system was able to create and utilize a topological map, identifying intersections and places of interest.
- Performance improves as the learned map becomes more useful through repeated task execution.
- The system successfully handled situations where items were not immediately visible, using strategies to search aisles and relevant areas.
Limitations and Future Work
- Error Handling: The current system assumes major errors in sensing will not happen. Integrating robustness to sensing errors and semantic-poor data is a future direction.
- Static Follower Procedure: The Follower currently uses a static procedure. The authors suggest it should use a similar representation to search plans for coping with contingencies and opportunistic behavior.
- Landmark Disambiguation: While the passive mapper attempts landmark disambiguation, the paper notes that it doesn't explicitly handle cases where a place of interest cannot be found.
- Dead Reckoning: The authors suggest that dead reckoning could be used to reject potential locations when sensor data alone is ambiguous, which is not currently implemented.
- Generalization to Other Domains: While tested in a grocery store, the system's ability to generalize to other structured environments would be a valuable extension.
Summary updated