BOMP: Bin-Optimized Motion Planning

In logistics, the ability to quickly compute and execute pick-and-place motions from bins is critical to increas- ing productivity. We present Bin-Optimized Motion Planning (BOMP), a motion planning framework that plans arm motions for a six-axis industrial robot with a long-nosed suction tool to remove boxes from deep bins. BOMP considers robot arm kinematics, actuation limits, the dimensions of a grasped box, and a varying height map of a bin environment to rapidly generate time-optimized, jerk-limited, and collision-free trajectories. The optimization is warm-started using a deep neural network trained offline in simulation with 25,000 scenes and corresponding trajectories. Experiments with 96 simulated and 15 physical environments suggest that BOMP generates collision-free trajectories that are up to 58 % faster than baseline sampling-based planners and up to 36 % faster than an industry- standard Up-Over-Down algorithm, which has an extremely low 15 % success rate in this context. BOMP also generates jerk-limited trajectories while baselines do not. Website: https: //sites.google.com/berkeley.edu/bomp. I. I N T RO D U C T I O N Robots are increasingly used for package handling and picking in logistics settings. When transporting thousands of packages each day, reductions in cycle time can significantly increase robot productivity. Particularly when working in deep bins, package handling robots are often equipped with long-nosed suction tools (e.g., the “bluction” tool from Huang, et al. [1]) to enable them to reach and manipulate packages throughout the deep bin. The long-nosed suction tool also allows the robot wrist and arm to remain far from obstacles and potential collisions. In deep cluttered bins, contents can shift after each pick, necessitating a strategy to rapidly compute pick-and-place motions using the latest sensor data (e.g., color and depth camera images). Practical approaches include heuristic planning, optimization-based motion planning, and sampling- based motion planning. A common heuristic trajectory, Up-Over-Down, lifts the package to clear all obstacles, moves horizontally over obstacles to the target location, then lowers. This is easy to implement and has negligible compute time; however, the motion is longer than necessary and often fails when using a long-nosed suction tool and planning in a deep bin. In particular, vertically lifting from the bottom to the top of the bin may be kinematically infeasible in deep bin environments. Optimization-based methods formulate and solve an opti- mization problem to find the best or fastest trajectory that avoids collisions. Sampling-based methods randomly sample 1The AUTOLab at UC Berkeley (automation.berkeley.edu) 2Carnegie Mellon University, Pittsburgh, PA, USA Fig. 1: Bin-optimized motion planning. BOMP executing a time-optimized, jerk-limited, collision-free trajectory moving a box from a bin to a drop-off point. We use the long-nosed “bluction” tool from Huang, et al. [1] to enable the robot to reach all parts of the deep bin, and an overhead RGBD camera to detect obstacles and target boxes. BOMP uses an optimization-based motion planner to compute the pick-and-place trajectory. In order to speed up the computation, a neural network warm-starts the optimizer. It accepts the obstacle environment, grasped box, and pick and place poses as input, and outputs an initial trajectory. and connect collision-free waypoints to find a path. The latter two methods yield significantly more successful and faster motions than Up-Over-Down, but at the expense of longer compute times. In prior work, we presented GOMP [2], an optimization- based motion planner incorporating time-optimization, ob- stacle avoidance, and grasp-optimization. GOMP computes fast motions between pick and place poses, and the grasp- optimization further speeds up motions by allowing pick and place poses to be optimized while retaining the same parallel-jaw grasp contact points. Subsequently, DJ-GOMP [3] further reduces compute time by using a neural network to warm-start motion planning for time-optimized and jerk- limited trajectories. DJ-GOMP is trained over a distribution of start and end points, assuming a fixed collision environment. However, in warehouse settings, boxes often move between picking actions. To address changing obstacle environments, we propose Bin-Optimized Motion Planning (BOMP). BOMP finds a time- optimized trajectory while considering collisions between the robot, a grasped box, and the obstacle environment. We integrate BOMP into an end-to-end bin-picking pipeline, which takes as input an RGBD image of the bin and outputs time-optimal trajectories (Figure 1). BOMP modifies and extends DJ-GOMP by adding the dimensions of a grasped box and a height map of the environment as inputs to the warm-start neural network, which enables adapting the 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) October 14-18, 2024. Abu Dhabi, UAE 979-8-3503-7769-9/24/$31.00 ©2024 IEEE 11055