Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Refinery: Active Fine-tuning and Deployment-time Optimization for Contact-Rich Policies

Bingjie Tang, Iretiayo Akinola, Jie Xu, Bowen Wen, Dieter Fox, Gaurav S. Sukhatme, Fabio Ramos, Abhishek Gupta, Yashraj Narang

TL;DR

Refinery addresses the gap between simulation success and industrial reliability in contact-rich robotic assembly by reducing performance variance across initial conditions. It combines Bayesian Optimization-guided fine-tuning to focus on high-uncertainty initial states with deployment-time Gaussian Mixture Model sampling to select favorable initializations, improving both individual policy robustness and sequence-level assembly success. In simulation, Refinery yields a mean improvement of 10.98 percentage points on 2-part assemblies (to $91.51\%$) and, when coupled with GMM deployment, achieves up to $96.35\%$ across multiple tasks, while real-world tests demonstrate strong per-step reliability and substantial sequence gains with some remaining challenges due to grasp stability and tolerance mismatches. The results highlight the complementary roles of uncertainty-aware fine-tuning and success-driven deployment for achieving industrial-grade, long-horizon robotic assembly.

Abstract

Simulation-based learning has enabled policies for precise, contact-rich tasks (e.g., robotic assembly) to reach high success rates (~80%) under high levels of observation noise and control error. Although such performance may be sufficient for research applications, it falls short of industry standards and makes policy chaining exceptionally brittle. A key limitation is the high variance in individual policy performance across diverse initial conditions. We introduce Refinery, an effective framework that bridges this performance gap, robustifying policy performance across initial conditions. We propose Bayesian Optimization-guided fine-tuning to improve individual policies, and Gaussian Mixture Model-based sampling during deployment to select initializations that maximize execution success. Using Refinery, we improve mean success rates by 10.98% over state-of-the-art methods in simulation-based learning for robotic assembly, reaching 91.51% in simulation and comparable performance in the real world. Furthermore, we demonstrate that these fine-tuned policies can be chained to accomplish long-horizon, multi-part assembly$\unicode{x2013}$successfully assembling up to 8 parts without requiring explicit multi-step training.

Refinery: Active Fine-tuning and Deployment-time Optimization for Contact-Rich Policies

TL;DR

Refinery addresses the gap between simulation success and industrial reliability in contact-rich robotic assembly by reducing performance variance across initial conditions. It combines Bayesian Optimization-guided fine-tuning to focus on high-uncertainty initial states with deployment-time Gaussian Mixture Model sampling to select favorable initializations, improving both individual policy robustness and sequence-level assembly success. In simulation, Refinery yields a mean improvement of 10.98 percentage points on 2-part assemblies (to ) and, when coupled with GMM deployment, achieves up to across multiple tasks, while real-world tests demonstrate strong per-step reliability and substantial sequence gains with some remaining challenges due to grasp stability and tolerance mismatches. The results highlight the complementary roles of uncertainty-aware fine-tuning and success-driven deployment for achieving industrial-grade, long-horizon robotic assembly.

Abstract

Simulation-based learning has enabled policies for precise, contact-rich tasks (e.g., robotic assembly) to reach high success rates (~80%) under high levels of observation noise and control error. Although such performance may be sufficient for research applications, it falls short of industry standards and makes policy chaining exceptionally brittle. A key limitation is the high variance in individual policy performance across diverse initial conditions. We introduce Refinery, an effective framework that bridges this performance gap, robustifying policy performance across initial conditions. We propose Bayesian Optimization-guided fine-tuning to improve individual policies, and Gaussian Mixture Model-based sampling during deployment to select initializations that maximize execution success. Using Refinery, we improve mean success rates by 10.98% over state-of-the-art methods in simulation-based learning for robotic assembly, reaching 91.51% in simulation and comparable performance in the real world. Furthermore, we demonstrate that these fine-tuned policies can be chained to accomplish long-horizon, multi-part assemblysuccessfully assembling up to 8 parts without requiring explicit multi-step training.

Paper Structure

This paper contains 16 sections, 6 equations, 5 figures, 3 tables, 2 algorithms.

Table of Contents

  1. INTRODUCTION
  2. RELATED WORK
  3. Robotic Assembly
  4. Improving Subtask Policies for Long-Horizon Tasks
  5. Problem Description
  6. Method
  7. Bayesian Optimization-Guided Fine-Tuning
  8. GMM Sampling for Deployment-Time Optimization
  9. Experimental Results
  10. Simulation-based Evaluation
  11. BO-Guided Fine-Tuning
  12. GMM-based Sampling for Deployment-Time Optimization
  13. Full Assembly Sequence
  14. Real-world Evaluation
  15. Limitations
  16. ...and 1 more sections

Figures (5)

  • Figure 2: Multi-part Assembly Dataset. We provide 5 multi-part assemblies derived from tian2023asap.
  • Figure 3: Comparison of success rate by using different acquisition function for fine-tuning in 5 multi-part assemblies (see \ref{['fig:multipart_dataset']} for asset IDs). X-axis corresponds to step index within each assembly.
  • Figure 4: Average success rates of baseline policies (AutoMate) and fine-tuned policies across 100 2-part assembly tasks. Policies are sorted by baseline success rate. BO-guided fine-tuning significantly improves the average success rate from 80.53% to 91.51%, with greatest improvements occurring for the most challenging tasks.
  • Figure 5: Comparison of success rate under different fine-tuning and deployment-time sampling strategies in 5 multi-part assemblies.
  • Figure 6: Real-world experiment setup. The Franka Research 3 (FR3) robot and a table-mounted Schunk EGK40 gripper is used to perform multi-part assembly tasks. 3D-printed 10000, 12099, and 15654 are shown in disassembled and assembled configurations.