Reconfiguring Procedural Knowledge for Compositional Robot Skill Adaptation
Organizations: Sungkyunkwan University · University of Texas at Austin · Samsung Electronics
Abstract
General-purpose behavior models increasingly internalize broad procedural knowledge during training, making their adaptation to new goals and skill interfaces less a problem of learning entirely new behaviors than of reconfiguring existing competence. We study this capability as reconfigurability, the ability to realize new behaviors by cost-efficiently restructuring existing components and relations under sparse governing information, where task specifications provide high-level goals and procedure specifications expose a limited set of reusable factors, leaving the concrete bindings and compositions required for execution underdetermined. We propose RecPo, a framework that represents skills task-agnostically through multiple procedure views and materializes executable skill sequences by composing stored transitions to realize task-specific changes. RecPo maintains a procedure knowledge base of skill-induced transitions, with each view exposing a reusable factor and preserving local skill connectivity. At deployment, it interprets task specifications through the same views and resolves the resulting compositional queries over the stored relations to bind skills for execution. On long-horizon tasks in an extended Franka Kitchen benchmark, RecPo achieves 91.9% ordered completion with 1.0 ms of task inference per episode across varying procedure support and task-record sparsity. Real-robot studies further demonstrate adaptation through local memory edits without task-specific retraining.
Figures & tables
| Appendix |
Appendix figures & tables31 assets
Supplementary material from the paper’s appendix.
Appendix
| Evidence | Statistic | Used in task materialization | Ablation axis |
| effect | decision points, active views | – | |
| admissible set , decision-point refinement | – | ||
| selection of , decision-point refinement | – | ||
| identification | , | decision points, selection of , projection | view selection |
| order | , | candidate support , path objective , decision-point refinement | – |
| † | final skill order | chaining prior |
| Stage | Hyperparameter | Value |
| Procedure Knowledge Base | DBSCAN neighborhood radius | |
| DBSCAN minimum sample count | ||
| salience nearest neighbors | ||
| Task materialization | support threshold | |
| refinement tolerance | ||
| beam width |
| ID | MW | KT | BB | TB | LS | SC | HC | ID | MW | KT | BB | TB | LS | SC | HC |
| Factor | RB | BB | TB | LS | SC | HC | MW | KT |
| Dimensions | – | – | – | – | – | – |
| Step | |||||||
| Subtask | MW | KT | BB | TB | LS | SC | HC |
| Compute resources | Clustering and numerical libraries | ||||
| CPU | AMD Ryzen Threadripper PRO 9975WX (32-core) | numpy | 2.2.6 | cuml | 25.12.0 |
| GPU | 2 NVIDIA GeForce RTX 5090 | scipy | 1.16.3 | cuvs | 25.12.0 |
| OS | Ubuntu 24.04.3 LTS | scikit-learn | 1.7.2 | cugraph | 25.12.2 |
| CUDA | 12.8 | umap-learn | 0.5.11 | cupy | 14.0.1 |
| Driver | 570.211.01 | pynndescent | 0.6.0 | ||
| Notation | Meaning |
| , , | Skill-context diversity, waypoint gap, and view resolution. |
| Number of evaluation seeds, each with episodes. | |
| Number of listed configurations. | |
| Total quality-evaluation episodes, , with . | |
| An axis averaged over its listed configurations. Additional averaging for Mean is stated in each caption. | |
| – | Not applicable. |
| Views | ||||
| configurations total |
| Experiment Config | ||||
| Views † | ||||
| configurations total | ||||
| Experiment Config | ||||
| Views † | ||||
| configurations total | ||||
| Experiment Config | ||||
| Views | ||||
| configuration | ||||
| configurations † | ||||
| configurations † | ||||
| configuration | ||||
| Experiment Config | ||||
| Views † | ||||
| configurations total | ||||
| Experiment Config | ||||
| Views † | ||||
| – | configurations total | – | ||
| Experiment Config | ||||
| Views † | ||||
| configurations total | ||||
| Experiment Config | |||||
| Views † | |||||
| Experiment Config | |||||
| Views † | |||||
| Experiment Config | ||||
| Views † | ||||
| configurations total | ||||
| Experiment Config | ||||
| Views † | ||||
| configurations total | ||||
| Experiment Config | ||||
| Views | ||||
| One subset per | ||||
| Experiment Config | |||||
| Report | Views | ||||
| One subset per | |||||
| One subset per | |||||
| Experiment Config | ||||
| Views | ||||
| Experiment Config | ||||
| Views † | ||||
| configurations total | ||||
| Experiment Config | ||||
| Task input | Views † | |||
| Dense | configurations total | |||
| Experiment Config | ||||
| Task input | Views | |||
| Dense | configuration | |||
| Dense | configurations † | |||
| Dense | configurations † | |||
| Dense | configuration | |||
| Experiment Config | ||||
| Views | ||||
| Dense | RecPo : S-DTW: – | |||
| Experiment Config | ||||
| Views | ||||
| RecPo : Others: – | ||||
| Experiment Config | ||||
| Views | ||||
| – | ||||
Robot Policy Adaptation
Robot policy adaptation changes or steers learned robot behaviors to handle new tasks, environments, hardware, and user preferences.
See all 194 papers