Robots are increasingly expected to perform delicate assembly tasks with incredible precision, often requiring accuracy within fractions of a millimeter. A newly published research paper introduces Facet-0, a robotic foundation model designed to improve how robots handle contact-rich manipulation tasks where precise positioning and sensitive force control are critical. This advancement could help robots better manage complex manufacturing processes that demand both spatial precision and safe, compliant interaction with objects.
Key Takeaways
- Facet-0 integrates multiple data types—vision, language, and physical forces—to predict the outcomes of its actions during contact-based tasks.
- The model uses a novel training approach combining multimodal learning and reinforcement learning to optimize both motion and contact forces.
- When tested on sub-millimeter assembly tasks, Facet-0 achieved an 82% success rate, far surpassing previous methods that reached only 15%.
- The system can adapt to different parts and environments with low latency, maintaining placement accuracy within 0.5 millimeters and responding within 50 milliseconds.
Robotic assembly at very fine tolerances is challenging because robots must not only move precisely but also carefully manage the forces they apply during contact with parts. Traditional robots often struggle with unexpected contact failures or minute misalignments that cause errors. Facet-0 addresses this by predicting the consequences of its actions on both movement and force levels, allowing it to plan motions that avoid damaging contacts or misplacements.
The researchers developed Facet-0 by unifying multimodal representation learning and reinforcement learning. In simpler terms, the robot learns from diverse inputs—visual data, language cues describing the task, and measurements of forces and torques (collectively called “wrenches”) experienced by its wrist. This combined data helps the model understand not just where to move but how the applied forces will affect the outcome. The training process uses “flow matching,” a technique that teaches the model to generate sequences of actions alongside predicted future force profiles.
To further improve decision-making, the system employs a special “Action-Wrench Critic” that evaluates how different motions, which might look similar in terms of task progress, actually differ in their contact outcomes. This helps the robot focus on critical interactions that determine success or failure. Additionally, the model includes a “bounded actor” component that allows quick adaptation to new parts or setups by reusing learned representations without retraining the entire system. This makes it practical for real-world manufacturing environments where conditions can change frequently.
The team trained Facet-0 on a large dataset called ManuFacet-1K, which includes 1,000 hours of synchronized force and motion data from multiple robot embodiments and manufacturing cells. This extensive training enabled the model to generalize across different tasks and hardware configurations.
With its high success rate and precise control, Facet-0 represents a significant step toward robots that can reliably perform complex, delicate assembly tasks in manufacturing. While further testing and refinement are needed before widespread deployment, this research opens new possibilities for automation in industries requiring fine manipulation and robust contact handling. Future work may explore extending the model to even more diverse tasks and integrating it with broader robotic systems for end-to-end manufacturing solutions.
Based on research published on arXiv by Haoyuan Deng, Haichao Liu, Wenkai Guo et al..