Organizations: School of Automation Science and Electrical Engineering, Beihang University · Beijing Institute for General Artificial Intelligence (BIGAI) · Delta Intelligence · State Key Laboratory of General Artificial Intelligence, School of Intelligence Science and Technology, Peking University
We devise a dual-arm cooperative planning framework for wiping table surface in the presence of scattered objects. Unlike conventional surface-cleaning strategies that only cover free areas and detour around obstacles, our framework explicitly assigns one arm to rearrange objects while the other performs coverage wiping, thereby enabling access to previously obstructed regions. Specifically, we first construct a set of wiping and object-rearrangement action primitives. Based on these primitives, we introduce Geo-MILP, a geometry-aware task planner formulated as Mixed-Integer Linear Programming (MILP). By integrating Logic-Based Benders Decomposition (LBBD), Geo-MILP iteratively optimizes action sequencingand timing to resolve geometric conflicts, promoting efficient cooperation between wiping and object rearrangement. We then employ K-Order Markov Optimization (KOMO) to generate smooth, collision-free motions to ensure feasible execution. Simulation and physical experiments demonstrate that the proposed framework not only increases reachable coverage by enabling access to obstructed areas, but also significantly improves time efficiency compared to Task and Motion Planning (TAMP) baselines, reducing planning time by up to 9 times and execution makespan by approximately 40%.
Figures & tables
Fig. 1 : A tpyical table wiping task. (a) Classic CPP: The robot detours around obstacles, leaving occluded regions uncleaned. (b) TAMP: The robot strictly alternates between rearrangement and wiping, leading to inefficient execution. (c) Ours (Geo-MILP based TAMP): The planner generates a synchronized schedule where wiping and rearrangement occur concurrently, achieving complete coverage with minimal makespan.
Fig. 2 : The proposed hierarchical framework. (a) CPP generates surface coverage waypoints and identifies obstructed nodes. (b) The task planner uses lbbd to jointly schedule concurrent bimanual actions. (c) The KOMO-based motion planner translates the discrete schedule into continuous, collision-free joint trajectories.
Fig. 3 : the Geo-MILP Task Planning framework. The framework employs a lbbd scheme to resolve the spatial temporal coupling between the two arms. The Master Problem (MILP) optimizes the high-level task schedule based on logical dependencies, while the Geometric Sub-Problem iteratively verifies the feasibility of candidate schedules and injects logic-based cuts to resolve geometric conflicts. This closed-loop process ensures the generated plan is both time-efficient and physically executable.
Fig. 4 : Experimental setups for the dual-arm tabletop wiping task. (a) The simulation environment used for evaluation. The semi-transparent overlays conceptualize the random spatial variance of the objects. For visual clarity, only a representative subset of the 10 randomized initializations is shown. (b) Real-world deployment on a dual Franka Research robotic system. (c) Hardware transferability demonstration on the Kuavo humanoid robot platform.
Fig. 5 : Qualitative execution sequence. (1) & (4): Our framework enables true bimanual concurrency: the left arm picks and removes the obstacle, allowing the right arm to wipe the previously occluded surface, followed by placing the object back. (2) & (3): Comparison of before and after states across different objects. Unlike static single-arm CPP detouring, our reversible relocation strategy achieves complete workspace coverage without leaving uncleaned patches.
Fig. 6 : Quantitative performance comparison. (a) Success Rate : Task success rates across varying clutter complexities. The shared legend applies to all subfigures. (b) Planning Time : Distributions highlighting the tight variance of Geo-MILP compared to search-based baselines. (c) Execution Makespan : Total execution time, showing significant bimanual efficiency gains.
∣K∣
Geo iters
MILP time (s)
KOMO time (s)
1
2.5±0.5
2.57±0.83
11.30±1.45
2
3.8±1.0
6.39±2.28
11.79±0.60
3
5.3±1.1
19.56±1.87
15.43±1.25
TABLE I : Geo-MILP statistics and planning-time breakdown. The results are reported as mean ± standard deviation over 10 randomized trials per complexity level.