Researchers have released PackLab, a framework for developing, training, and evaluating multimodal large language models (MLLMs) on robotic bin packing, according to a paper posted on Hugging Face. The task is described as long-horizon sequential decision-making, since every object placement changes the free space available for the objects that follow.
The authors note that most existing approaches rely on hand-crafted geometric heuristics tuned to predefined objectives, or on reinforcement learning policies trained through trial and error under fixed configurations. While MLLMs have recently been applied to packing, the paper argues their potential for closed-loop sequential decisions across heterogeneous packing setups remains underexplored.
PackLab is presented as three components. PackLab-Suite is a physics-based simulation platform that generates diverse packing trajectories at scale and evaluates their physical outcomes. PackLab-VLM is a packing-specialized MLLM that reads evolving object and container states to jointly select objects and predict placements in a closed loop. PackLab-Bench offers standardized packing scenarios at multiple difficulty levels for systematic evaluation.
In experiments, the team reports that PackLab-VLM outperforms conventional packing heuristics, traditional reinforcement learning methods, and general-purpose MLLMs on average across object sets and container configurations. The authors frame this as evidence of the potential of MLLMs for long-horizon robotic packing.
Code, model, dataset, and benchmark are released publicly: the code is on GitHub, while the PackLab-VLM-9B model and the PackData-20K and PackLab-Bench datasets are hosted on Hugging Face.