It Takes Two to Tangram: Multi-Agent Object Centric Reinforcement Learning for Assembly
Authors: Paul Knaußt, Kay Pompetzki, Nico Bohlinger, Niklas Funk, Cristiana de Farias, Rolf Findeisen, Jan Peters
Organizations: Intelligent Autonomous Systems Lab, Technical University of Darmstadt, Germany. · Control and Cyber-Physical Systems Laboratory, Technical University of Darmstadt · Hessian.AI, Germany. · German Research Center for AI (DFKI), SAIROL, Germany. · Robotics Institute Germany (RIG).
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/
Figures & tables
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 configurations. We provide simulation-ready models, together with a robot-compatible physical counterpart consisting of 3D-printed pieces and an insertion tool for planar manipulation.
Fig. 4: Policy ablations. On the top row, we show the architecture ablation on TangramBench-Trio. Here we report task success, fraction of pieces within 5 cm of their targets, position and yaw errors, and contact rate throughout training. On the bottom, the effect of the reward formulation. We ablate the shared team reward and inter-piece contact cost.
Fig. 5: Success comparison on assemblies containing a different piece count.
Method
Succ.
Err.
Contact
Time
Compute
(%)
(cm)
(%)
(s)
TangramBench-Cluster
GOMA (ours)
75
1.0
4
8
0.68 ms
APF
3
3.8
99
80
≈0
RRT-J
73
1.9
44
20
774 ms
RRT-J +replan
98
0.3
12
21
49 ms
TABLE I: GOMA against the APF controller and joint (RRT-J) and per-piece (RRT-P) RRT planning, planned once or replanned every 5 s, on the held-out TangramBench-Cluster and TangramBench-Puzzle.