Facet-0: A Robotic Foundation Model for Contact-Rich Precise Manipulation

· Editorial Team estimated
contact-rich-manipulation robot-foundation-models precision-assembly force-control reinforcement-learning

Facet-0 is a robotic foundation model that predicts and values the contact consequences of its actions: it aligns a causal wrench history with vision-language semantics and kinematic state, uses flow matching to generate action chunks jointly with the wrist-wrench profile they are expected to induce, and concentrates reinforcement learning on decisive interactions via a distributional Action-Wrench Critic and contact-selective credit. Trained on ManuFacet-1K, a 1,000-hour force-synchronized corpus spanning three embodiments, it reaches 82% mean success on five sub-millimeter computer-assembly tasks versus 15% for the strongest baseline, with 0.5 mm placement accuracy and 50 ms command latency.

Paper · arXiv:2609.01596

Sub-millimeter assembly is one of the hardest thresholds in industrial robotics: it demands spatial precision, compliant interaction, and robustness to contact failures all at once. Classical solutions lean on precision fixtures and teach programming, treating contact as an disturbance to be avoided, and they rarely generalize across parts. Facet-0 starts from a different premise: a genuinely useful assembly foundation model must learn to predict and value the contact consequences of its actions — not just where the hand should go, but what force that motion will produce, whether it will jam, and whether it will succeed.

Core Innovation

The architecture is organized around a joint action-wrench proposal:

  • Joint proposal and flow-matching generation — a causal wrist-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, so decisions are made with their contact outcomes in mind.
  • Distributional Action-Wrench Critic — trained on deployment rollouts, it distinguishes motions with similar task progress but very different contact outcomes, adding contact-level resolution to value estimation.
  • Contact-selective credit assignment — phase-aware rewards and contact-selective credit concentrate policy improvement on the interactions that actually decide success, rather than diluting it across trivial motion phases.
  • Lightweight bounded adaptation — a small bounded actor reuses the frozen representation for fast on-robot adaptation; RL stays defined over executable Cartesian actions while an auxiliary wrench head preserves predictive, non-commanded action-contact coupling.

Results

  • Data: ManuFacet-1K — a 1,000-hour force-synchronized corpus spanning three embodiments and multiple manufacturing cells, among the largest assembly force-interaction datasets publicly described.
  • Success rate: 82% mean success across five sub-millimeter computer-assembly tasks, versus 15% for the strongest baseline — a decisive gap.
  • Precision and latency: 0.5 mm placement accuracy and 50 ms command latency, compatible with real-time closed-loop cell deployment.

Limitations

The evaluation is confined to five computer-assembly tasks, so cross-domain generalization is unproven; each part class still requires task-level adaptation through the bounded actor rather than being fully plug-and-play; the distributional critic depends on deployment-rollout data, implying non-trivial cold-start costs; and the abstract does not detail the distribution of failure modes.

Industry Implications

Consumer-electronics and computer assembly are among the highest-value automation niches in manufacturing, and connector insertion, shield placement, and flex-cable routing have long resisted automation. An 82%-versus-15% jump moves sub-millimeter assembly from “not viable” to “production-ready,” while 50 ms latency and online adaptation target exactly the frequent changeovers of real manufacturing cells. For equipment vendors, the combination of a force-synchronized corpus, a foundation model, and lightweight adaptation may reset the cost structure of precision assembly: away from part-specific fixtures and teaching, toward data- and model-centric rapid deployment.