An experimental robotic system cleared all 30 bins in tests on three automotive parts, but the result came with a clear trade-off. The model-free baseline cleared every bin, yet its grasp-success average was far below the model-based baseline. The hybrid system combined a model-based backbone with fallback grasp generation and online learning. Its later levels were intended to take over when the first approach reached a deadlock, meaning the current plan could no longer make progress.
The document is a preprint, labeled arXiv version 1 and dated 28 August 2026. In the baseline comparison, bin clearance averaged 50.9% for the model-based setup and 100.0% for the model-free setup across all three parts. Grasp success told the opposite story: 99.1% for model-based versus 45.6% for model-free overall.
Bin clearance is the share of the starting parts removed from a bin. Grasp success is the share of grasp attempts that succeeded. That distinction matters because the model-based baseline stopped at deadlocks and did not execute grasps in the hardest configurations, so its grasp-success average is not a perfectly like-for-like comparison with the model-free result. The report gives descriptive averages, does not report formal hypothesis tests, and does not attach uncertainty intervals to the table averages.
A fallback built into the machine
At Level I, the system used a model-based object detector, a 6D pose estimator for an object's position and orientation, and a grasp database. The later levels addressed three potential perception failures. Levels II, III and IV used Contact-GraspNet, trained on GraspClutter6D, to generate model-free fallback grasps.
The baselines covered parts A, B and C with three bins per part, starting with 50 parts per load carrier. Hybrid experiments used five bins per part and ran all four hierarchy levels with online self-learning and Wilson-score ranking.
During operation, electrical-gripper width readings were used to identify whether a part slipped out. For each grasp point, the system updated its count of attempts and successful attempts, then ranked grasps using the upper border of a Wilson score interval. For the 95% confidence interval, it set the Wilson parameter z to 1.96.
Where the extra layers mattered
The aggregate numbers conceal sharp differences. Model-based clearance averaged 33.3% for part A, 34.7% for part B and 84.7% for part C. Model-free clearance was 100.0% for each. For grasp success, model-based averaged 96.1% for A and 100.0% for both B and C, while model-free averaged 71.4%, 29.8% and 55.0%, respectively.
In hybrid tests, bin clearance was 100% across all 30 bins. The manually initialized condition covered 15 bins; the report records 18 deadlocks overcome at Level II, one at Level III and three at Level IV. Most recorded resolutions occurred at Level II, with smaller numbers at the higher fallback levels.
Manual-Init hybrid grasp success averaged 91.9% for A, 83.6% for B and 89.3% for C. Those figures were below the model-based baseline averages reported for the same parts, even as the hybrid cleared every tested bin. Manual initialization took between 30 and 45 minutes per part.
With Empty Init, the authors report improvement over the model-free baseline for all three parts, but final performance varied by object. The reported hypothetical pre-deadlock grasp-success penalties were 2.6% for A, 1.2% for B and 0.0% for C. The analysis also notes persistent grasp-depth issues for part B.
A promising test with narrow boundaries
Fallback processing added 1.01 seconds for grasp generation and 1.36 seconds for collision checking. The collision check examined 4,096 predicted grasps with 180-degree replications and cost about 0.3 milliseconds per grasp. The grasp database was extended by 22 grasps for A, 10 for B and 12 for C, while the reported effect on Level I cycle time was negligible.
The findings are narrow. They come from three automotive part types and one stated hardware configuration, so they do not establish performance across broader part catalogs, bins, robot embodiments or factories. The study also does not establish long-term learning behavior or the best parameter choices; the supplied analysis identifies extended experiments and ablations as still needed.
Nor do the experiments demonstrate fully zero-shot plug-and-produce operation. Rare cases involving entangled parts were manually corrected, and foreign objects or extra packaging could make Level IV unsuitable. The authors point to CAD-based initialization and improved model-free grasp generators as open directions.
Within that scope, the authors interpret the results as evidence that model-free grasp generation can complement a model-based backbone, resolve deadlocks and expand the grasp database during operation. The work was supported by the Federal Ministry for Economic Affairs and Energy based on a resolution of the German Bundestag with DARP. As a preprint, the report remains an early technical result rather than a broad validation of factory deployment.
Paper data and sources
Original title: Picking Bins Empty: A Hierarchical Hybrid Approach with Online Self-Learning of Grasp Points for Reliable Industrial Bin-Picking
Authors: Florian Töper, Samarth Kishor Yelvande, Jan Niklas Ewertz et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text