Facet-0: A Robotic Foundation Model for Contact-Rich Precise Manipulation
2026-09-01 • Robotics
RoboticsMachine Learning
AI summaryⓘ
The authors created Facet-0, a robot system designed to handle very precise assembly tasks that require careful contact and motion control. It learns from a large dataset how actions affect forces and combines vision, language, and motion data to predict outcomes and improve its actions through reinforcement learning. Facet-0 adapts to new parts by reusing learned knowledge while fine-tuning some parts for specific tasks. Their system performed much better than existing methods on tiny assembly tasks, achieving high accuracy and success rates. This shows how combining multiple types of information and learning methods can improve robot precision in real-world settings.
robotic assemblyreinforcement learningmultimodal representationcontact forceswrench predictionflow matchingaction-criticCartesian actionsfine manipulationdata-driven robotics
Authors
Haoyuan Deng, Haichao Liu, Wenkai Guo, Yuan Ling, Zaijia Yang, Yuanjiang Xue, Haosheng Sun, Liangzi Wang, Ziwei Wang
Abstract
Real-world robotic assembly at sub-millimeter tolerances demands spatial precision, compliant interaction, and robustness to contact failures. We present Facet-0, a robotic foundation model that predicts and values the contact consequences of its actions. Facet-0 unifies multimodal representation learning and reinforcement learning (RL) post-training around a joint action-wrench proposal: a causal wrench history is aligned with vision-language semantics and kinematic state, and flow matching generates each action chunk together with the future wrist-wrench profile it is expected to induce. Deployment rollouts train a distributional Action-Wrench Critic to distinguish motions with similar task progress but different contact outcomes, while phase-aware rewards and contact-selective credit concentrate policy improvement on decisive interactions. To accommodate part-specific dynamics, a lightweight bounded actor reuses the frozen representation for on-robot adaptation; RL remains defined over executable Cartesian actions, while an auxiliary wrench head preserves predictive, non-commanded action-contact coupling. Trained on ManuFacet-1K, a 1,000-hour force-synchronized corpus spanning three embodiments and multiple manufacturing cells, the bounded task-adapted system reaches 82% mean success on five sub-millimeter computer-assembly tasks, compared with 15% for the strongest baseline, with 0.5 mm placement accuracy and 50 ms command latency.