Intelligent Grasping of Unknown Waste Objects for Enhanced Recycling and Reuse

Project Idea

Despite Denmark’s well-established waste separation systems, only a small fraction of certain waste streams—particularly plastic—is cleanly reused or recycled. Much of it is either incinerated or downcycled. To maximize value, waste must be sorted into purer fractions (e.g., by plastic type such as PP or PE), or ideally, reusable items (e.g., standard transport boxes) should be identified and extracted before recycling.

Traditional sorting methods rely on shredding items into flakes, which are then separated. However, this approach is inefficient and limits reuse potential. A more effective strategy is to identify and grasp whole objects (e.g., bottles, jugs, electronics) directly from the waste stream, enabling higher-value reuse and reducing processing complexity.

This task is challenging because waste objects are often unknown and lack predefined models (e.g., CAD). Grasping such objects requires AI systems that can infer grasp strategies from sensor data (e.g., camera images) without prior knowledge of the object.

Our previous work showed that off-the-shelf models like GraspNet-1Billion and SuctionNet-1Billion do not generalize well to real-world waste scenarios. We therefore trained a custom model for a limited set of plastic objects using a single suction-based gripper. However, this approach was labor-intensive, and some objects could not be grasped with that gripper type.

We aim to address three core challenges:
- C1: Different objects require different grippers which are currently not part of our system.
- C2: Large-scale grasp data must be collected autonomously with minimal human intervention.
- C3: AI models must be developed to select both the appropriate gripper and grasp location based on sensor input.

This project will advance robotic waste handling by enabling adaptive grasping of unknown objects. The system will be applicable to general waste streams and easily adaptable to specialized contexts (e.g., hospital waste, corporate take-back programs), supporting more efficient and sustainable reuse and recycling.

Method

To address C1, we will expand the system with multiple gripper types. In addition to the existing suction cup, we will develop a novel high-flow, low-vacuum gripper inspired by household vacuum cleaners and evaluate other gripper designs for suitability in waste handling.

To tackle C2, we will integrate a circular conveyor system that enables the robot to repeatedly grasp and drop objects, allowing the collection of thousands of grasp trials with minimal manual effort.

With the capabilities from C1 and the data from C2, we will address C3 by developing and evaluating AI models for gripper selection and grasp planning. We will begin with simple models and incrementally increase complexity.

Depending on progress, we will explore advanced capabilities such as: (a) Incremental retraining and continuous learning to adapt to new object types. (b) Detection of high-value items (e.g., mobile phones) for prioritized extraction. (c) Optimization strategies based on parameters such as object value, processing sequence, or runtime efficiency.

The project is funded by Fabrikant Vilhelm Pedersen og Hustrus Legat.