Introduction
The CoBot (Collaborative Robot) project at Carnegie Mellon University represents a significant milestone in long-term autonomous service robotics. Over several years of deployment, a fleet of CoBots has successfully navigated more than 1,000 km within multi-floor office buildings. Unlike robots designed for isolated laboratory settings, CoBots are engineered to handle the inherent messiness and unpredictability of real-world human environments. This robustness is achieved through a combination of innovative localization techniques, intelligent multi-robot scheduling, and a unique "symbiotic" relationship with human occupants.
Challenges in Human Environments
Autonomous service robots typically face two primary hurdles: environmental dynamism and physical constraints.
- Dynamic Environments: Traditional localization algorithms often rely on static maps. In active offices, however, furniture is moved, boxes are stacked, and people are constantly in motion. These "short-term features" (STFs) introduce noise that can cause standard Markov-based localization to fail.
- Physical Limitations: Performing complex manipulation tasks, such as pressing elevator buttons or handling mail, requires expensive and heavy robotic arms. Building a fully independent robot capable of these tasks is often cost-prohibitive and technically daunting.
The CoBot project addresses these challenges by moving away from the ideal of total independence, focusing instead on a "symbiotic autonomy" model where the robot is robust enough to navigate autonomously but intelligent enough to request human assistance for tasks beyond its physical capabilities.
Technical Pillars
Episodic non-Markov Localization (enML)
To handle dynamic environments, the researchers developed Episodic non-Markov Localization (enML). Traditional Markov localization assumes that the current state depends only on the immediate previous state. In contrast, enML utilizes a "Varying Graphical Network" (VGN) to track three distinct feature types: Long-Term (permanent structures like walls), Short-Term (movable furniture), and Dynamic (moving people).
By treating navigation as a sequence of "episodes," the robot can maintain internal consistency regarding unmapped objects while discarding stale data once it reaches a point of independence. This approach allows the robot to treat moved furniture as temporary landmarks rather than sensor noise, significantly improving reliability in changing spaces.
Multi-Robot Scheduling with Transfers
Managing a fleet requires sophisticated coordination. The CoBot scheduler supports time and location constraints and introduces the concept of transfers. For example, one robot can retrieve a package on an upper floor and hand it to another robot at the elevator. This allows robots to specialize in certain areas or floors, optimizing the overall efficiency of the fleet. The system employs Mixed Integer Programming (MIP) for small task sets and Traveling Salesperson Problem (TSP) approximations for larger scales.
Symbiotic Autonomy
Symbiotic autonomy is a planning framework where "asking for help" is treated as a primary action. When a CoBot needs to change floors, it navigates to the elevator and proactively asks a nearby human to press the button. The planner uses a decision-theoretic model to select the best person to ask, considering factors like distance, the person's current activity, and how frequently they have been interrupted previously.
Evaluation and Results
The system was evaluated through the "1,000km Challenge" at CMU’s Gates-Hillman Center. Key findings include:
- Localization Robustness: The use of enML increased the mean distance between human interventions from 4.79 km to 8.13 km. In challenging open areas like atria, the improvement was even more significant, jumping from 0.62 km to 4.42 km.
- Operational Efficiency: Proactively seeking human help for elevator operation reduced task completion times by an average of 4 minutes compared to passive waiting.
- User Integration: During a two-week trial, over 100 occupants successfully used the robots for tasks such as snack delivery, mail transport, and visitor escorting.
Conclusion
The CoBot project demonstrates that perfect independence is not a prerequisite for effective service robotics. By combining robust localization that accounts for environmental change with a symbiotic approach to human-robot interaction, the CoBots have proven their ability to provide reliable service in complex, everyday environments. Future research will focus on learning human preferences to minimize social friction and improving natural language interfaces for more intuitive interaction.
Summary updated