Venue: CoRL 2025 (Best Paper Award, one of two) · Authors: Yunsheng Tian, Joshua Jacob, Yijiang Huang, et al. (13 authors) — MIT CSAIL, Autodesk Research, ETH Zurich, Texas A&M; senior author Wojciech Matusik (MIT) Category: Dexterous / Bimanual Manipulation Trend tag: Integrated planning + learning for contact-rich dual-arm arXiv: 2506.05168

flowchart TB
  CAD[Multi-part<br/>object CAD] --> P[Hierarchical planner<br/>precedence → sequence →<br/>grasp → motion<br/>+ auto fixture gen]
  P --> SK[Lightweight RL<br/>generalist contact-rich policy<br/>equivariance + residual on plan]
  SK --> EXEC[Two-arm execution<br/>zero-shot sim-to-real]

Robotic assembly research is usually a single benchmark object. Real assembly involves many parts, contact-rich insertions, and coordination between two arms — a problem where pure learning lacks structure and pure planning lacks dexterity.

Fabrica integrates planning and learning into an end-to-end dual-arm pipeline:

• Hierarchical task-and-motion planning for long-horizon assembly: a hierarchy of precedence → sequence → grasp → motion planning, with automated fixture generation so assembly is feasible on any dual-arm robot. The planner has a parallelizable design and is optimized for control stability.
• Lightweight RL for contact-rich steps: rather than per-task skills, a single RL framework trains generalist contact-rich policies that generalize across object geometries, assembly directions, and grasp poses. Learning is guided by equivariance and by residual actions obtained from the plan (the plan provides a nominal trajectory; the policy learns a residual correction).
• The trained policies transfer zero-shot to the real world — no domain knowledge or human demonstrations are required.

The core contribution is the first complete and generalizable real-world multi-part assembly system without domain knowledge or demonstrations, generalizing across object families (furniture, toys, industrial equipment) rather than a single benchmark object.

• CoRL 2025 Best Paper Award (one of two). Demonstrates end-to-end assembly of diverse multi-part objects that neither planning nor learning alone could handle.
• ~80% successful steps in real-world dual-arm execution on a benchmark suite of multi-part assemblies resembling industrial and daily objects (furniture, toys, industrial equipment).
• Policies trained on diverse assemblies transfer to novel assemblies with performance comparable to specialist policies trained per-object, and transfer zero-shot from simulation to the real world.
• The authors note remaining limitations toward industrial-grade robustness for such long-horizon, precise tasks.

A marker that the "planning + learning" integration — which had fallen out of fashion vs. end-to-end policies — still wins on truly contact-rich, combinatorial tasks. Complements the end-to-end VLA thread by showing where structure still pays.

• arXiv: https://arxiv.org/abs/2506.05168
• Project page: https://fabrica.csail.mit.edu/
• Code: https://github.com/yunshengtian/Fabrica
• OpenReview: https://openreview.net/forum?id=aSUNzvEJIf
• CoRL 2025 awards page: https://2025.corl.org/program/awards

• Survey (CoRL 2025)

← Back to CoRL-2025