DBPF: A Framework for Efficient and Robust Dynamic Bin-Picking

TL;DR

DBPF tackles dynamic bin-picking by enabling reactivity to moving objects and dynamic obstacles through a closed-loop pipeline that integrates TAMN-guided suction pose generation, velocity-matched approach, Kalman-filter motion prediction, horizon-based trajectory optimization, and a resight policy. The framework couples a pose-selection metric M(s) with a TAM score to robustly select suction poses, while SDF-based collision modeling and MPPI-based planning ensure safe, real-time trajectories. Empirical results show DBPF achieving an average SR of about 84% in fully dynamic scenarios with zero collisions and outperforming static SPA and other baselines, demonstrating improved efficiency and reliability in dynamic bin-picking. The approach holds promise for industrial automation by reducing time overhead and increasing adaptability under dynamics, with future work focusing on grasping with a two-finger gripper, handling rotational motions, and learnable environment representations.

Abstract

Efficiency and reliability are critical in robotic bin-picking as they directly impact the productivity of automated industrial processes. However, traditional approaches, demanding static objects and fixed collisions, lead to deployment limitations, operational inefficiencies, and process unreliability. This paper introduces a Dynamic Bin-Picking Framework (DBPF) that challenges traditional static assumptions. The DBPF endows the robot with the reactivity to pick multiple moving arbitrary objects while avoiding dynamic obstacles, such as the moving bin. Combined with scene-level pose generation, the proposed pose selection metric leverages the Tendency-Aware Manipulability Network optimizing suction pose determination. Heuristic task-specific designs like velocity-matching, dynamic obstacle avoidance, and the resight policy, enhance the picking success rate and reliability. Empirical experiments demonstrate the importance of these components. Our method achieves an average 84% success rate, surpassing the 60% of the most comparable baseline, crucially, with zero collisions. Further evaluations under diverse dynamic scenarios showcase DBPF's robust performance in dynamic bin-picking. Results suggest that our framework offers a promising solution for efficient and reliable robotic bin-picking under dynamics.

Figures (5)

• Figure 1: Dynamic bin-picking: The robot picks up a single object among a cluster of arbitrary objects stacked in a moving bin on a belt conveyor. Our novel framework enables the robot to achieve dynamic bin-picking, ensuring collision-free and adaptive motion. Our method outperforms traditional bin-picking methods in terms of time efficiency, showcasing exceptional effectiveness and reliability in dynamic scenarios.
• Figure 2: Framework overview: The DBPF is implemented in a fully closed-loop manner to facilitate the robot's reactivity in dynamic bin-picking. An eye-in-hand object camera captures the point cloud of moving objects within the bin at 10Hz. With the suction pose generation and selection, a set of candidate suction poses $S$ are generated, and an optimal pose $s_{opt}$ (yellow cylinder) is selected considering factors like the Tendency-Aware Manipulability (through TAMN), pose consistency, height preference, and suction score. Motion prediction forecasts bin's displacement $\vec{r}$ to advance the $s_{opt}$ as $s_{tar}$ and offers bin velocity $v_{bin}$. An eye-to-hand environment camera obtains RGB-D images at 30Hz, and the dynamic obstacle perception module detects the pose of the moving bin $p_{bin}$. Two Signed Distance Functions $SDF_{static}$ and $SDF_{dynamic}$ are maintained actively corresponding to environmental collisions. Horizon-based discrete trajectory optimization solves the optimal trajectory $\zeta_{opt}$ with objectives like pose-matching $f_{goal}$ and velocity-matching $f_{vel}$. Dynamic obstacle avoidance is realized via collision constraints. The first point $q_1$ of each optimal path $\zeta_{opt}$ is executed at 50Hz, and the robot's current joint position $q_t$ is constantly updated. Lastly, a task-level planning model integrates these modules tightly to effectively achieve dynamic bin-picking while preventing "Poor Observations" through the resight policy.
• Figure 3: Task-level planning model consists of six actions for dynamic bin-picking. The actions are the Wait action in a standby state, the Observe action to perceive the suction pose and bin state, the Track action to follow the target pose, the Surpass action for rushing forward and regaining a decent view for observation, the Approach action to travel the final distance to contact with object surface, and the Pick & Place action to attach the object from the bin and lift, and drop at placing location.
• Figure 4: Snapshots of dynamic bin-picking. A: (a)-(f) demonstrate the process of picking an object from a clutter of objects from a moving bin on a belt conveyor. B: (a)-(f) depict Neat and Clutter arrangements of objects in varying bins ranging from Small to Large in size.
• Figure 5: Success rate (left) and total time (right) of dynamic bin-picking. *_FD: Methods under the Fully Dynamic scene. *_DS: Methods under the Disturbed Static scene.