This paper introduces a novel framework for dual-arm robot tabletop rearrangement that moves away from rigid, synchronous "handoffs." By combining Movable Stacking for efficient buffering and Gap-Aware Push Transfers for flexible cross-boundary object exchange, the system achieves significantly higher efficiency in cluttered environments compared to traditional methods.
TL;DR
The researchers developed a dual-arm planning framework that replaces fixed, synchronous object handoffs with asynchronous "push transfers" across a shared boundary. By integrating this with movable stacking—where stacks of objects are treated as single units—the system reduces coordination bottlenecks and execution costs, particularly as tabletop clutter increases.
What Problem Does This Paper Address?
Rearranging objects on a crowded table is a classic "combinatorial" challenge in robotics. When you add a second arm to improve speed and coverage, you introduce a massive coordination headache.
Traditionally, if Arm A needs to give an object to Arm B, they meet at a fixed handoff configuration—a specific point in space where they both must be present at the same time. This creates a "synchronization bottleneck": one arm often has to wait for the other, and the fixed location might be blocked by other clutter. In dense scenes, these rigid constraints lead to deadlocks or highly inefficient, roundabout paths.
Methods
The authors propose a framework built on three main pillars:
- Joint Monte Carlo Tree Search (MCTS): The "brain" of the system is a joint MCTS planner. It explores a tree of possible actions (move, stack, push, wait) for both arms simultaneously, balancing the need to make immediate progress toward the goal with the need to explore new configurations.
- Movable Stacking: Instead of treating the table as a flat surface where objects can only be placed in empty spots, the system can stack objects. Crucially, these stacks are "movable"—if the bottom object is moved, the entire stack moves with it. This effectively turns a stack into a single high-density "buffer" unit, freeing up precious table space.
- Gap-Aware Push Transfer: This is the core innovation for inter-arm coordination. Instead of a fixed meeting point, the planner scans the boundary between the two arms' workspaces to find "gaps"—collision-free intervals. One arm performs a pre-push alignment, pushes the object across the boundary, and then goes back to its own tasks. The second arm can then retrieve the object whenever it is ready. This asynchronous approach decouples the timing of the two arms.
Datasets
The researchers evaluated their framework using simulated tabletop environments:
- Workspace: A 1m x 2m tabletop.
- Objects: 12 heterogeneous object classes (boxes, pots, bowls, fruit, cutlery) with different support properties to test realistic stacking.
- Scenarios: 20 random scene pairs (initial vs. goal configurations) for each object count $N$, where $N$ ranged from 5 to 12 objects to simulate increasing levels of clutter.
Key Insights and Ideas
- Asynchrony is Efficiency: By allowing one arm to "drop off" an object via a push and move on, the system avoids the "wait-and-see" idle time inherent in traditional handoffs.
- Dynamic Boundary Selection: Treating the crossing location as a planning variable (finding the best "gap") rather than a fixed design choice allows the robot to adapt to the specific geometry of the clutter.
- Density-Dependent Advantage: The more crowded the table, the more the "fixed handoff" becomes a liability. The proposed method's flexibility shines brightest when space is at a premium.
Key Results
The framework was benchmarked against a state-of-the-art "Fixed Handoff" baseline. The results showed a clear trend:
- In Light Clutter ($N=5, 6$): Both methods performed similarly, with the fixed handoff sometimes having a slightly lower execution cost because the paths are simple.
- In Heavy Clutter ($N=12$): The proposed method reduced the number of Joint Steps by approximately 10.3% and the overall Execution Cost by about 5.1%.
- Scalability: As the number of objects increased, the gap between the proposed method and the baseline widened, proving that dynamic transfer selection is essential for complex rearrangement tasks.
Limitations and Future Work
While the results are promising, the paper leaves some areas for future exploration:
- Search vs. Optimization: The MCTS planner is "progress-driven" (it wants to reach the goal) rather than strictly "cost-minimizing." This means it might find a valid plan that isn't the absolute shortest possible path.
- Real-World Physics: The experiments were conducted in simulation. Transitioning to real-world hardware would require handling the uncertainties of friction and contact during the "push" phase more robustly.
- Action Complexity: The current primitive set is discrete. Future work could integrate more complex non-prehensile maneuvers (like sliding or pivoting) directly into the search loop.
Summary updated