Learning to Build: Autonomous Robotic Assembly of Stable Structures Without Predefined Plans

Jingwen Wang; Johannes Kirschner; Paul Rolland; Luis Salamanca; Stefana Parascho

cs.ROFeb 27, 2026

Learning to Build: Autonomous Robotic Assembly of Stable Structures Without Predefined Plans

Authors: Jingwen Wang, Johannes Kirschner, Paul Rolland, Luis Salamanca, Stefana Parascho

Organizations: Lab of Creative Computation, ´Ecole Polytechnique F´ed´erale de Lausanne · Swiss Data Science Center

Abstract

This paper presents a novel autonomous robotic assembly framework for constructing stable structures without relying on predefined architectural blueprints. Instead of following fixed plans, construction tasks are defined through targets and obstacles, allowing the system to adapt more flexibly to environmental uncertainty and variations during the building process. A reinforcement learning (RL) policy, trained using deep Q-learning with successor features, serves as the decision-making component. As a proof of concept, we evaluate the approach on a benchmark of 15 2D robotic assembly tasks of discrete block construction. Experiments using a real-world closed-loop robotic setup demonstrate the feasibility of the method and its ability to handle construction noise. The results suggest that our framework offers a promising direction for more adaptable and robust robotic construction in real-world environments.

Figures & tables

Fig. 1: Left: Problem Definition: Example of a construction task, defined by the construction space, target locations, obstacle regions, and available unit blocks. Right: Tasks: A set of 15 tasks used to evaluate the reinforcement learning approach.

Fig. 2: Example episode for Task 8. The successor-feature images visualize the discounted future block occupancy predicted for the selected action, providing a spatial representation of the policy’s anticipated construction. Reward and successor-feature values are dimensionless.

Fig. 3: Closed-loop robotic assembly workflow. Given the current constructed structure (state $S$ ) and the task specification, the trained policy selects the next action $A$ for the robot to execute. After the robot places the block in the physical world, ArUco-based pose estimation is used to detect the updated block configuration. This information is then fed back into the simulation to update the state $S$ for the next decision step.

Fig. 4: Closed-loop robotic assembly setup: Left: Real-world construction scene with the robotic arm. Right Top: 3D-printed blocks labeled with ArUco markers for visual tracking. Right Bottom: Custom L-shaped suction gripper used for block manipulation.

Fig. 5: Training performance across five independent random seeds. The plots show the mean number of solved tasks, average cumulative reward, and average assembly size across the 15 training tasks; shaded regions indicate one standard deviation across runs. The policies rapidly improve during the first 10 episodes and subsequently maintain similar performance, solving approximately 13–14 tasks on average. Meanwhile, the average assembly size decreases from about 10 to 5–6 blocks, indicating that the policies learn more material-efficient construction strategies.

ID	P	A*
1	352	579
2	471	1212
3	386	1725
4	604	433
5	552	1143
6	317	445

TABLE I: Node visitation counts for the trained policies and A* across all tasks. ID : Task number; P : Nodes visited by the trained policy; A* : Nodes visited by A* algorithm. Task 9 with A* did not succeed, marked by † . Total nodes visited: D = 5763, A* = 24866 (D/A* = 0.23).

Fig. 6: Generalization performance under four types of unseen task variations. (a) Horizontal translation of the entire task, used to evaluate robustness to spatial translation. (b) Symmetry breaking by shifting target locations relative to the obstacles. (c) Interpolation between pairs of training-task geometries, producing unseen intermediate layouts. (d) Local perturbations of target and obstacle positions. The curves report mean task success rate as the magnitude of each variation increases.

Fig. 7: Successfully completed tasks in real-world construction. The top row shows the simulated assemblies, while the bottom row shows the corresponding real-world constructions using closed-loop feedback. Notably, in Task 3 & 12, the policy produces assemblies that differ from the simulation.

Fig. 8: Unsuccessful and untested tasks in real-world construction. Task 4 failed due to robotic infeasibility—specifically, a gripper collision that prevented block placement. Task 7 failed due to limited policy robustness in the presence of accumulated construction noise. Tasks 9 and 11 failed due to structural instability under minor physical perturbations, despite being classified as stable in simulation. Task 13 was not physically tested because its simulated assembly was judged to be marginally stable and difficult to realize reliably in the real-world setup.

ID	S	Avg	Max
1	1/1	0.04	0.08
2	1/1	0.01	0.03
3	1/1	0.03	0.08
4	0/4	0.03	0.06
5	1/1	0.03	0.08
6	1/1	0.01	0.03

TABLE II: Construction results for all tasks. ID : Task number; S : Successful trials / Total trials; Avg : Average relative placement offset; Max : Maximum relative placement offset. Offsets are normalized by block size. Note that we did not test Task 13.

Explore similar work

Learning to build covering structures with continuous adjustments
Sep 8, 2026Gabriel Vallat, Maryam Kamgarpour, Stefana ParaschoScalable Robot Learning Efficient Exploration
Skill Sequence Planning for Collaborative Multi-Robot Construction
Sep 22, 2026Xi Wang, Bo Fu, Carol C. Menassa +2Construction Collision-Free Trajectories
Department of Construction Science, Texas A&M University, College Station, TX 77843, USA · Amazon Robotics, North Reading, MA 01864, USA · Department of Civil and Environmental Engineering, University of Michigan, MI 48109, USA +1
BrickCraft: Visuomotor Skill Composition with Situated Manual Guidance for Long-Horizon Interlocking Brick Assembly
May 8, 2026Jichuan Yu, Bowei Li, Zhenran Tang +4Geometric Guidance Craft
Tsinghua University · Carnegie Mellon University

Sep 8, 2026, cs.ROJ/K move · Enter open · S save

Learning to build covering structures with continuous adjustments

Gabriel Vallat, Maryam Kamgarpour, Stefana Parascho

Scalable Robot Learning Efficient Exploration

Open paper AskPDF arXiv

Robotic construction offers the potential to use materials more efficiently and create complex geometries, but current methods rely on rigid, high-precision plans that cannot accommodate the tolerances, inaccuracies, and unexpected changes inherent in physical fabrication. In this work, we introduce a reinforcement learning approach that forgoes predefined plans entirely, instead generating construction sequences adaptively as the structure is built. Our method operates on graph-structured state representations and a mixed (parameterized) action space, requiring both discrete block selection and continuous placement parameters. Because the stability simulation of a structure is computationally heavy, we develop an efficient exploration strategy by incorporating unilateral edges into graph neural networks, extending soft actor-critic (SAC) to this hybrid setting. We evaluate our algorithm, HSAC, against the prior method hybrid-PPO (HPPO), demonstrating significantly higher asymptotic performance and good sample efficiency. We also demonstrate HSAC's robustness to hyperparameter choices and its exploration capability, handling up to 10 discrete actions without performance degradation. Finally, we validate our approach on a physical two-robot setup, successfully building a spanning arch with 3D-printed blocks in closed-loop execution, confirming that policies trained in simulation transfer to real hardware.

Abstract

Explore similar work

Learning to build covering structures with continuous adjustments

Skill Sequence Planning for Collaborative Multi-Robot Construction

BrickCraft: Visuomotor Skill Composition with Situated Manual Guidance for Long-Horizon Interlocking Brick Assembly