Qubit-to-Motion: Traceable Action Provenance in Hybrid QRL for Robotic Bin-Packing Sequencing
Organizations: Zhejiang University Robotics Institute, Ningbo, China · Department of Electrical and Computer Engineering, University of New Brunswick, Fredericton, NB, Canada
Abstract
Hybrid quantum reinforcement learning has been explored for robotic decision-making, but evaluations often emphasize task success, return, or convergence, leaving unresolved whether quantum circuit outputs influence constrained actions and remain traceable through robot execution. We present Qubit-to-Motion, an evidence-centered framework for support-constrained sequencing and simulated robotic placement. Four Pauli- readouts provide the sole learned input to the actor and critic, while a support mask enforces admissible object selection. Each decision is recorded and linked via step indices and SHA-256 fingerprints to a sealed route, deterministic Cartesian and inverse-kinematics waypoints, prescribed joint motion, and MuJoCo contact replay. A frozen checkpoint is evaluated across five deterministic 30-object scene variants. All 150 placements satisfy support, containment, and settling criteria. A fixed-state zero-readout intervention changes of decisions, demonstrating dependence on the joint circuit representation. Mean macro travel is ~m, inference requires ms per route, and replay achieves m mean final position error with 0.356mm maximum interpenetration. These results establish reproducible decision-to-contact traceability. The implementation is available at linkpink{the anonymous repository}.
Figures & tables
| Setting | Value |
|---|---|
| Training seed | 2025 |
| Demonstration scenes / decisions | 18 / 540 |
| BC epochs / learning rate | 15 / |
| PPO updates / decisions per update | 2 / 120 |
| Optimization passes per PPO update | 3 |
| Configured PPO batch size | 128 |
| Scene seed | (m) | (m) | (%) | (%) | Position mean ( m) | Position max. ( m) | Passed / total |
|---|---|---|---|---|---|---|---|
| 41 | 34.499 | 34.348 | 90.0 | 93.1 | 8.883 | 16.170 | 30 / 30 |
| 73 | 34.804 | 34.536 | 90.0 | 93.1 | 8.544 | 16.158 | 30 / 30 |
| 109 | 34.678 | 34.572 | 90.0 | 93.1 | 9.050 | 16.174 | 30 / 30 |
| 151 | 34.283 | 34.437 | 90.0 | 93.1 | 8.544 | 16.156 | 30 / 30 |
| 197 | 34.421 | 34.293 | 90.0 | 93.1 | 8.796 | 16.165 | 30 / 30 |
| Readout substitution | (%) | (%) | Mean |
|---|---|---|---|
| 90.0 | 93.1 | 0.102 | |
| 57.3 | 59.3 | 0.136 | |
| 50.7 | 52.4 | 0.109 | |
| 69.3 | 71.7 | 0.135 | |
| 36.0 | 37.2 | 0.081 | |
| Reverse component order | 45.3 | 46.9 | 0.070 |
Robotic RL
Robotic reinforcement learning uses trial-and-error learning to help robots control, navigate, and manipulate objects.
See all 323 papers