TL;DR
This paper introduces a system that allows modular robots to autonomously build structures to overcome environmental obstacles, enabling them to complete tasks that would otherwise be impossible. By combining novel perception algorithms for identifying augmentable features with a high-level planner that decides when and where to build, these robots can dynamically adapt their environment to achieve their goals.
What Problem Does This Paper Address?
Modular robots, while capable of self-reconfiguration, often struggle with traversing large obstacles. This limitation restricts the scope of tasks they can perform. Existing work has focused on either the robots' ability to reconfigure themselves or on collective construction of structures. However, there's a gap in enabling modular robots to autonomously build structures in their environment specifically to overcome extrinsic, high-level task challenges, much like a human might use a ladder to reach a high shelf. This paper addresses this by enabling robots to perceive their environment, identify opportunities for augmentation, and then build structures to facilitate task completion.
Methods
The core of the system involves three main components:
- Hardware: The system utilizes the SMORES-EP modular robot platform, augmented with novel passive "wedge" and "block" modules. These modules are designed to be carried and traversed by the SMORES-EP robot and are equipped with AprilTags for localization.
- Perception (Environment Characterization): A novel probabilistic, template-based algorithm is used to identify "augmentable features" in the environment. This algorithm takes a 2.5D elevation map and searches for specific feature templates (e.g., a "ledge") that can be augmented. It identifies candidate locations and orientations where structures could be beneficial.
- Planning (High-Level Planner): This planner, building on prior work, now integrates environment augmentation as a strategy. It synthesizes robot controllers (finite state automata) from user-defined task specifications. Crucially, it can now decide not only to reconfigure the robot but also to augment the environment by building a structure if necessary to complete a task. It uses the output from the environment characterization to inform these decisions.
Datasets
Not discussed in the paper. The experiments were conducted in realistic office environments using custom-built augmentation modules and the SMORES-EP robot.
Key Insights and Ideas
- Environment Augmentation as a Task Strategy: The most significant insight is framing environment augmentation not just as a construction task, but as a strategic tool for achieving higher-level goals. This moves beyond robots building for construction's sake to robots building to enable other tasks.
- Bridging Perception and Action for Augmentation: The paper effectively links perception (identifying where to build) with planning (deciding if and how to build) and action (physically building and using the structure). This closed-loop system is key to autonomous operation.
- Overcoming Modular Robot Limitations: The proposed system directly addresses the inherent weakness of modular robots in traversing large obstacles by providing a mechanism to create custom-fit solutions (ramps, bridges) on demand.
- Generalizability: The framework is designed to be general, allowing for different feature templates and structure types, suggesting potential for broader applications beyond the specific experiments shown.
Key Results
The system was validated in two hardware experiments:
- Experiment I (Desk Drawers): The robot successfully identified a "ledge" feature created by an opened drawer and autonomously acquired and placed a ramp to climb onto the drawer to access a second, higher drawer. In a variation where the "ledge" was absent (empty drawer), the system correctly identified that augmentation was not possible and the mission could not be completed.
- Experiment II (Gap Crossing): The robot identified a 16 cm gap as a "gap" feature and autonomously assembled and placed a bridge using available wedge and block modules to cross to a target zone.
The paper reports on 8 runs for each experiment, with success rates of 25% for Experiment I and 37.5% for Experiment II. Failures were attributed to hardware issues, perception errors (mis-detections or inaccurate localization), and navigation failures (cumulative SLAM errors).
Limitations and Future Work
- Robustness: The authors acknowledge that robustness remains a significant challenge, with hardware failures and perception/navigation errors being the primary causes of failure.
- Speed and Scale: The current work focuses on demonstrating the concept with small structures (three elements) and does not prioritize construction speed or the ability to build very large structures.
- Automated Assembly Planning: The assembly plans for structures were created manually. Automating this process is a key area for future work.
- Parallel Construction: Exploring multi-robot construction for faster assembly is suggested.
- Material Transport: Developing mechanisms for transporting building materials to task locations is identified as an open challenge for scaling up.
- Generalization to Other Platforms: While the concepts are general, direct application to other platforms would require adapting sensing and perception capabilities.
Summary updated