This paper introduces Facet‑0, a robotic foundation model that predicts and values the contact consequences of its actions. The goal is to give robots the spatial precision and safe touch needed for real‑world assembly tasks that require sub‑millimeter fits. Facet‑0 combines visual and language cues, the robot’s motion state, and a history of forces to decide and score actions that will produce desirable contact outcomes.

At a high level, the system learns a joint representation of what the robot sees, what it has been doing, and what forces (or torques) it has felt. The authors use a “wrench” history—wrench meaning the forces and torques measured at the robot’s wrist—and align that history with vision and kinematic state. A flow‑matching model then generates short action chunks along with a prediction of the future wrist‑wrench over that chunk. During deployment, rollouts are used to train an Action‑Wrench Critic, which is a value model that prefers motions that lead to good contact outcomes even when those motions look similar in terms of task progress.

The paper also describes practical steps for on‑robot use. To adapt to part‑specific dynamics, they freeze the learned representation and add a small, bounded actor that is trained on the robot. Reinforcement learning still outputs executable Cartesian actions (motions in 3D space), while an auxiliary wrench head keeps predicting the action–contact coupling even when that prediction is not commanded. They call this combination a robotic foundation model because the frozen representation can be reused and adapted to specific tasks.

Training used a new dataset called ManuFacet‑1K: a 1,000‑hour force‑synchronized corpus collected across three robot embodiments and multiple manufacturing cells. With this training and the bounded task adaptation, the system achieved a mean success rate of 82% on five sub‑millimeter computer‑assembly tasks. For comparison, the strongest baseline reached 15%. The system also reported 0.5 mm placement accuracy and a 50 ms command latency, indicating both precision and fast control.