Facet-0 Enables Robots to See, Insert Precisely, and Recover from Errors in Precise Assembly

Precise assembly in manufacturing demands sub‑millimeter contact execution and autonomous error recovery, beyond visual perception alone.

Facet-0, a robot foundation model, was evaluated on five real computer‑assembly tasks (RAM, CPU, Disk, GPU, etc.). It achieved an average task success rate of 82% , compared with 10% for π0.5 and 4% for GR00T N1.7. Placement precision reached 0.5 mm with an instruction latency of about 50 ms . The evaluation comprised 23 sub‑goals and assembly gaps of 0.10–0.30 mm , highlighting the difficulty of the final few millimetres.

The training pipeline contributed to the performance gain: pre‑training yielded 16% success, reinforcement‑learning (RL) post‑training raised it to 38% , and lightweight adaptation further increased it to the final 82% . After deployment, human‑intervention rate dropped from 47% to 24% and failure‑recovery rate rose from 44% to 81% .

ManuFacet‑1K, the supporting dataset, contains roughly 1000 hours of force‑synchronized demonstrations collected on UR7e, xArm, and Franka arms for RAM, CPU, GPU, and Disk installations. Each frame records multi‑view images, language commands, end‑effector pose, gripper state, and six‑dimensional force/torque, annotated with skill stages (approach, align, insert, press, seat, fasten, retreat). The dataset also retains failure traces, human takeovers, and recovery actions, enabling learning of both successful trajectories and contact failure modes.

Facet-0 aligns visual‑language‑action semantics with force feedback through a three‑step pipeline:

Ingest multi‑view images, language instructions, robot state, and force/torque history; force data enters the model as an independent signal.

A coarse action expert generates a force‑aware primitive; a fine‑grained expert refines it using real‑time force data.

A high‑frequency compliant controller executes the final command while enforcing workspace and safety constraints.

RL post‑training decomposes the binary “success/failure” signal into per‑step quality assessments. Credit is assigned to sub‑tasks (align, insert, press, seat); successful retreat and re‑alignment after a jam also earn positive credit. The value model combines motion and force/torque changes, rewarding gentle contacts over aggressive pushes. High‑value contact experiences are fed back to the action expert, forming a closed loop: run → anomaly detection → human correction/autonomous recovery → value evaluation → policy update.

This approach differs from naïve continual‑training pipelines by explicitly surfacing rare failure moments, evaluating them, and converting them into learnable recovery experience, thereby reducing human intervention and improving recovery rates while preserving sub‑millimeter accuracy.

In a zero‑shot test on an unseen memory module, training with only 10 demonstrations (~3 h) and updating 6.6% of the parameters achieved 45% success, compared with the strongest baseline at 5% .

Future work includes expanding long‑tail part and fixture coverage, exploring unified semantic‑force modeling, and pursuing safe continual learning for long‑term deployment.

Project page: https://pine-lab-ntu.github.io/facet-0/ Paper: https://arxiv.org/abs/2609.01596