Abstract
Grasping a novel target object in constrained environments (e.g., walls, bins, and shelves) requires intensive reasoning about grasp pose reachability to avoid collisions with the surrounding structures. Typical 6-DoF robotic grasping systems rely on the prior knowledge about the environment and intensive planning computation, which is ungeneralizable and inefficient. In contrast, we propose a novel Collision-Aware Reachability Predictor (CARP) for 6-DoF grasping systems. The CARP learns to estimate the collision-free probabilities for grasp poses and significantly improves grasping in challenging environments. The deep neural networks in our approach are trained fully by self-supervision in simulation. The experiments in both simulation and the real world show that our approach achieves more than 75% grasping rate on novel objects in various surrounding structures. The ablation study demonstrates the effectiveness of the CARP, which improves the 6-DoF grasping rate by 95.7%.
Introduction
Most robotic grasping research focuses on "table-top" scenarios—clean, open surfaces where a robot can easily approach an object from above using 3-DoF grasping. However, real-world environments like warehouse bins, refrigerators, or cluttered shelves are "constrained." In these settings, a simple top-down approach often fails because the robot's arm or gripper would hit a wall or a neighboring object.
Traditional solutions rely on heavy collision-checking algorithms that iteratively test thousands of possible paths, which is computationally expensive and often fails without a perfect geometric model of the surroundings. This research addresses target-driven grasping in constrained environments, focusing on quickly finding a grasp that is both stable on the object and safe from environmental collisions.
Methodology
The proposed system utilizes a two-part pipeline consisting of perception and grasping modules.
Perception Module
To identify specific target objects, the system employs an RGB-D Siamese Network. By incorporating a depth channel, the network distinguishes between objects with similar colors but different shapes more effectively than RGB-only models. This approach also helps bridge the "sim-to-real" gap, allowing models trained in simulation to perform well in the physical world.
Grasping Module
The core innovation is the Collision-Aware Reachability Predictor (CARP), which works alongside a Grasp Stability Predictor (GSP):
- GSP: A 3D CNN that analyzes the object's local geometry to predict grasp stability.
- CARP: A 3D CNN that evaluates a voxelized view of the surrounding structures relative to a proposed 6-DoF grasp pose. It outputs the probability of a collision-free path.
The final feasibility score is the product of the stability score and the collision-free probability. The robot executes the pose with the highest combined score.
Experimental Setup
The system is trained entirely via self-supervision in simulation using CoppeliaSim and the Bullet physics engine. The training data includes 60,000 labeled point clouds for CARP and 50,000 data points for GSP. Evaluation was conducted on both known and novel objects in simulation and on a real-world Franka Emika Panda robot.
Results and Discussion
The inclusion of CARP significantly enhances performance in challenging scenarios, such as when objects are tucked into corners.
Key findings include:
- Simulation Performance: In challenging arrangements, the proposed method achieved a 93.55% planning rate and an 80.65% grasping rate, vastly outperforming baselines that lacked collision awareness.
- Real-World Validation: On a physical Panda robot, the system maintained a 91.40% planning rate and an 80.65% grasping rate across various arrangements, including novel objects.
- Ablation Study: CARP improved the 6-DoF grasping rate by 95.7% compared to systems considering only kinematic reachability.
The success of the model stems from decoupling grasp stability ("can I hold it?") from reachability ("can I reach it?"). Voxelizing the environment relative to the gripper allows the model to learn general spatial patterns of safety regardless of specific bin shapes.
Conclusion
While the current system demonstrates high success rates, it is limited by a fixed voxel resolution (2.5cm³), which may struggle with fine-grained obstacles. Future work aims to incorporate various robot hand shapes to improve hardware generalization and to explore more complex, highly cluttered scenes.
Summary updated