Pack it in: Packing into Partially Filled Containers Through Contact

TL;DR

This work tackles packing an object into a partially filled container by leveraging contact-based interactions to rearrange existing contents and create space. It introduces a physics-informed trajectory optimization framework using iLQR within an MPC loop, a physics-aware perception system, and a high-level placement planner to propose feasible packing poses in clutter. The approach is validated through real-robot experiments and simulation, showing that PackItIn outperforms baselines by enabling successful insertions in challenging, occluded scenarios and highlighting key sim-to-real gaps such as perception accuracy and contact modeling. The results demonstrate the practical potential for automated, space-efficient bin packing in warehouses, with identified avenues for improving robustness and scalability in highly cluttered environments.

Abstract

The automation of warehouse operations is crucial for improving productivity and reducing human exposure to hazardous environments. One operation frequently performed in warehouses is bin-packing where items need to be placed into containers, either for delivery to a customer, or for temporary storage in the warehouse. Whilst prior bin-packing works have largely been focused on packing items into empty containers and have adopted collision-free strategies, it is often the case that containers will already be partially filled with items, often in suboptimal arrangements due to transportation about a warehouse. This paper presents a contact-aware packing approach that exploits purposeful interactions with previously placed objects to create free space and enable successful placement of new items. This is achieved by using a contact-based multi-object trajectory optimizer within a model predictive controller, integrated with a physics-aware perception system that estimates object poses even during inevitable occlusions, and a method that suggests physically-feasible locations to place the object inside the container.

Figures (5)

• Figure 1: Snapshots of the proposed packing through contact system operating on real robotic hardware. The system uses contact-based interactions between the inserted object and objects in the container to clear space and place it within the container. Please also see attached video.
• Figure 2: Control system architecture for MPC on real robotic hardware.
• Figure 3: Two visualisation of the same packing scene with different sampled packing poses. The green object is the one to be packed, all other objects are present in the container before packing. The left image is the first frame when the packed object is teleported inside the container and the right image is after the physics simulator has been integrated $M$ times. The top row shows the best packing pose found for this scene, as it has minimal disturbance to several objects surrounding it. The bottom row incurs a much larger cost from Eq. \ref{['eq:placement_cost']} as it is unable to displace the purple object inside the container plane and instead, causes the object to move upwards outside of the container, implying this would be an unsuitable target packing pose.
• Figure 4: Overhead views of 5 of the 40 packing scenes used during hardware experiments. Top row shows the state of the scene before packing and the bottom row shows the state of the scene after the object (tomato passata) was packed inside the container. Four of these trials were successful (green outline) and one was unsuccessful (red outline). The unsuccessful case was an object detached failure, hence the object is in a different orientation.
• Figure 5: Box plots of the metrics defined in Eqs. \ref{['eq:pos_error']}--\ref{['eq:per_error']} for PackItIn trials. Each box indicates the interquartile range (IQR), with whiskers extending to 1.5$\times$IQR. The maroon line denotes the median, while the blue triangle marks the mean.