It Takes Two to Tangram: Multi-Agent Object Centric Reinforcement Learning for Assembly

Paul Knaußt, Kay Pompetzki, Nico Bohlinger, Niklas Funk, Cristiana de Farias, Rolf Findeisen, Jan PetersarXiv 2026 · arXiv

TL;DR

The authors introduce GOMA (Graph networks for Object-centric Multi-agent Assembly), a reinforcement learning framework that treats individual assembly parts as cooperative agents. By combining Graph Neural Networks (GNNs) with Multi-Agent Proximal Policy Optimization (MAPPO), the system learns to coordinate complex, simultaneous object motions to solve Tangram puzzles without collisions or deadlocks.

What Problem Does This Paper Address?

In robotic assembly, knowing the final destination of a part is only half the battle. In tight, cluttered spaces—like fitting the last few pieces into a puzzle—the path to the goal is often blocked by other components. This creates a "spatiotemporal coordination" problem: which piece should move first, and how can they move together without getting stuck?

Traditional methods often rely on Task and Motion Planning (TAMP), which can be computationally expensive and struggle with the "narrow passages" of configuration space. This paper explores whether we can instead learn a reactive feedback policy that understands the spatial relationships between parts and coordinates their velocities in real-time.

Methods

The researchers propose GOMA, which leverages two main architectural pillars:

  1. Object-Centric Representation: Instead of viewing the entire scene as one giant state vector, the system treats each piece as a node in a graph. A GNN processes the pieces' current poses, target poses, and their relative distances (edges).
  2. Multi-Agent Reinforcement Learning (MARL): Each piece is treated as an agent. They all share a single policy (trained via MAPPO) that outputs translational and angular velocities. A centralized "critic" looks at the whole scene to evaluate how well the team is doing.
The GOMA architecture uses GNNs to encode relationships and a shared policy head to command each piece.

The reward function is a mix of individual progress (moving toward your own target) and team success (minimizing the maximum error in the scene and completing the whole assembly). This encourages pieces to "make room" for one another.

Datasets

To test GOMA, the authors introduced TangramBench, a procedural benchmark based on the classic seven-piece Tangram puzzle. It features three levels of increasing geometric difficulty:

  • Trio: Only three pieces; a simplified coordination task.
  • Cluster: Seven pieces in irregular, relatively open configurations.
  • Puzzle: Seven pieces in tightly packed, edge-to-edge configurations (the hardest level).
Fig. 2: TangramBench provides a graded benchmark for planar geometric assembly, ranging from the three-piece TangramBench-Trio to seven-piece TangramBench-Cluster and tightly constrained TangramBench-Puzzle configura- tions. We provide simulation-ready models, together with a robot-compatible physical counterpart consisting of 3D- printed pieces and an insertion tool for planar manipulation.

The benchmark includes over 3 million procedurally generated training configurations and a set of held-out evaluation tasks.

Key Insights and Ideas

  • Relational Structure is King: The biggest performance jump came from adding the GNN. While a standard MLP (Multi-Layer Perceptron) struggled to coordinate, the GNN allowed agents to "see" how their neighbors' positions restricted their own movement.
  • Decoupling Control from Coordination: By training the policy to output object-level velocities rather than low-level robot joint torques, the authors achieved impressive zero-shot sim-to-real transfer. The same policy trained in simulation worked on a physical Franka robot by simply having the robot move one piece at a time in a "round-robin" fashion.
  • Team Rewards vs. Individual Drive: Ablation studies showed that individual rewards are enough to get pieces near their goals, but the shared "team reward" is essential for the final, precise "click" into place that completes the assembly.
Ablation studies show that GNN-based relational reasoning is essential for high success rates.

Key Results

  • Success Rates: GOMA achieved ~90% success on the "Trio" tasks and 75% on "Cluster" tasks. On the most difficult "Puzzle" tasks, it reached 41% success.
  • Speed and Efficiency: Compared to joint RRT planning (a standard sampling-based planner), GOMA was significantly faster at inference (0.68ms vs. 49ms–1.1s for RRT) and resulted in much lower contact rates (4% vs. 12–37% for RRT).
  • Sim-to-Real: The authors successfully demonstrated the policy on a real-world setup using 3D-printed pieces and a custom peg-in-socket tool, proving the abstraction's robustness.
Fig. 1: GOMA solving the tangram puzzle on our TangramBench. Given the target assembly (right), we focus on the complex task of coordinating feasible object motions constrained to the support plane. GOMA predicts object motions that avoid deadlocks and collisions under varying clearance to achieve the final assembly.

Limitations and Future Work

The current work makes a few simplifying assumptions:

  • Known Targets: The system is told exactly which piece goes where; it doesn't solve the "shape matching" part of the Tangram puzzle.
  • 2D Constraints: The pieces are constrained to a flat surface. Extending this to 3D assembly (like IKEA furniture) would introduce much more complex contact physics.
  • Sequential Execution: While the policy supports concurrent motion, the physical demonstration used a single robot arm moving pieces one by one. Future work could involve bimanual or multi-robot setups to fully realize the "simultaneous" nature of the learned policy.

Summary updated

Abstract

Geometric assembly requires fine spatiotemporal coordination to place parts without causing collisions or deadlocks under tight clearances. To address this, we present an object-centric reinforcement learning framework that treats each part as an agent and learns a shared policy for continuous assembly control. A graph neural network represents the pieces, their target poses, and their spatial relationships, and predicts translational and angular velocities for each component. Per-agent proximal policy optimization with a centralized graph critic directly learns motion and temporal coordination within a single feedback policy. We study this formulation through planar Tangram assembly, where pieces must slide and rotate into a given configuration while remaining on the assembly surface. Procedural generation of target configurations and randomized initial arrangements support training and evaluation across diverse assembly problems. Our experiments isolate the roles of relational structure and multi-agent optimization, and compare learned control with sampling-based planning. We demonstrate concurrent assembly in simulation and transfer the policy to sequential execution on a real robot, as our object-centric formulation decouples learned assembly coordination from low-level physical control. Project page: https://tangrambench.org/
BibTeX
@misc{knaut2026it,
  author = {Paul Knaußt and Kay Pompetzki and Nico Bohlinger and Niklas Funk and Cristiana de Farias and Rolf Findeisen and Jan Peters},
  title = {{It Takes Two to Tangram: Multi-Agent Object Centric Reinforcement Learning for Assembly}},
  year = {2026},
  eprint = {2611.08814},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.08814},
}
Figure