Contact-Rich Robotic Manipulation in Construction via Zero-Shot Learning: A Diffusion Policy-Guided Adaptive Control

Roman Ibrahimov; Salma Mozaffari; Arash Adel

cs.ROAug 22, 2026

Contact-Rich Robotic Manipulation in Construction via Zero-Shot Learning: A Diffusion Policy-Guided Adaptive Control

Authors: Roman Ibrahimov, Salma Mozaffari, Arash Adel

Organizations: Princeton University · Princeton University, Princeton, NJ 08544, USA

Abstract

Construction robotics and automation offer promising means of improving productivity, alleviating workforce shortages, and reducing workers' exposure to physically demanding tasks. However, reliable contact-rich robotic assembly remains challenging under tight tolerances, fabrication inaccuracies, and uncertain contact dynamics. To address this challenge, we present a framework coupling diffusion policies trained on simulation-generated pose and force/torque data with an L1-inspired adaptive controller that corrects policy-predicted actions online to compensate for unmodeled contact dynamics. We benchmark the framework against baselines in timber joinery, pipe fitting, and sequential full-scale truss assembly. It achieves 100% success on single-task assemblies and 90-100% success across sequential truss assembly subtasks, with lower, more stable contact forces than the baselines. By enabling zero-shot sim-to-real transfer for force-aware contact-rich assembly, the framework reduces costly, labor-intensive real-world data collection for policy training and advances scalable, robust automation of multistage assembly, motivating extension to broader contact-rich manipulation tasks in construction.

Figures & tables

Figure 1: Methods overview: we generate synthetic data in simulation using the robot position at the flange center and force/torque feedback at the robot flange (left) to train a CNN-based diffusion policy (center), then, we develop an adaptive controller to correct the policy-predicted actions ( $\Delta x,\Delta y,\Delta z$ ) of the trained policy, facilitating a zero-shot policy transfer and improving performance in real-world deployment (right).

Figure 2: The six-axis UR20 arm and custom end effector.

Figure 3: Example of sampling for trajectory generation (a) horizontal starting offset in the $xy-$ plane with $r_{samp}=$ 5 mm, and (b) vertical starting offset and descending heights during the insertion phase with $z_{app}=$ 60 mm and $z_{max}=$ 20 mm.

Figure 4: Example of utilizing parallel simulation environments for large-scale data collection.

Figure 5: Adaptive controller architecture, including state predictor (Section 3.4.1 ), adaptation law (Section 3.4.2 ), and control law (Section 3.4.3 ).

Figure 6: Single contact-rich assembly tasks in construction: mortise and tenon timber joint assembly (left) and pipe fitting (right).

Figure 7: Mortise and tenon assembly in sim and real (left) and pipe fitting in sim and real (right).

Symbol	Description	Value
$L_{r}$	Learning rate	$2\mathrm{e}{-4}$
$W_{d}$	Weight decay rate	$1\mathrm{e}{-6}$
$B$	Batch size	1024
$E$	Number of epochs	1000
$T^{p}_{o}$	Observation horizon	2
$T_{p}$	Action prediction horizon	16

Table 1: Parameters used for policy training and execution for both tasks.

Symbol	Description	Timber joint	Pipe fitting
$a_{pred}$	Predictor bandwidth constant	-62.8	-62.8
$\Gamma$	Adaptation gain	5	80
$\lambda$	Decay rate	0.98	0.99
$k_{d}$	Force damping gain	0.005	0.005
$F_{0}$	Force activation threshold (N)	30	30
$\sigma_{\max}$	Max disturbance (mm)	1	1

Table 2: Parameters for our proposed adaptive controller.

Symbol	Description	Timber joint	Pipe fitting
M	Mass (kg)	700	75
K	Stiffness (kN/m)	500	0.5

Table 3: Parameters for the compliance controller adapted from 59 .

Figure 8: Full-scale assembly of a gable truss: the bottom chord contains one central mortise for the vertical post and two side mortises for the rafters; each rafter is also connected to the post through a lap joint. All dimensions are in millimeters.

Figure 9: Joint variations in the full-scale assembly of the gable truss: vertical mortise-and-tenon joint between the post and bottom chord (left), angled mortise-and-tenon joint between a rafter and the bottom chord (center), and lap joint between a rafter and the post (right).

Figure 10: The developed workcell for the full-scale assembly experiment with pick and gripping stations.

Figure 11: Truss assembly sequence from left to right and top to bottom. First, the bottom chord is placed, followed by the post. Once the post is inserted and released by the UR20, the ABB arm grasps the post to provide support during the subsequent rafter assemblies. The left and right rafters are then assembled sequentially. After the final rafter is inserted, both the post and rafter are released.

Symbol	Description	Truss assembly
$a_{pred}$	Predictor bandwidth constant	-62.8
$\Gamma$	Adaptation gain	1
$\lambda$	Decay rate	0.98
$k_{d}$	Force damping gain	0.001
$F_{0}$	Force activation threshold (N)	150
$\sigma_{\max}$	Max disturbance (mm)	1

Table 4: Parameters for our proposed adaptive controller for the full-scale gable assembly.

Task	Ours	Baseline	Compliant policy	Environment
Timber joint	100	87	95	Real
Timber joint	90	73	73	Sim
Pipe fitting	100	100	100	Real
Pipe fitting	73	66	69	Sim

Table 5: Average success rates (%) per execution mode per task in simulation and real-world.

Figure 12: Peak force distributions across the execution modes for both tasks in simulation and real-world settings.

			$F_{max}$ (N)		One-Sided Mann-Whitney U Test
Condition	Method	$N$	$\mu$	$s$	$U$	$p_{\text{corr}}$	SL	CLES	Cohen’s $d$
Timber joint Real	Ours	40	152.5	5.0	—	—	—	—	—
	Baseline	40	159.6	20.0	654.0	0.161	ns	0.59	0.477
	Compliant policy	40	166.0	10.6	107.0	$<$ 0.001		0.93	1.601
Timber joint Sim	Ours	200	101.5	60.5	—	—	—	—	—
	Baseline	200	197.1	147.0	10485.5	$<$ 0.001		0.74	0.849

Table 6: Statistical comparison of peak force magnitude

F_{max}

(N) across methods.

p_{corr}

is the p-value after Bonferroni correction with significance levels (SL) defined as: *** =

p<0.001

, ** =

p<0.01

, * =

p<0.05

, and ns = not significant.

Joint Name	Success rate (%)	Environment
Post	100	Real
Left Rafter	90	Real
Right Rafter	100	Real

Table 7: Success rates for 10 rollouts per task in real-world.

Contact-Rich Robotic Manipulation

Contact-rich robotic manipulation studies how robots plan and control physical interactions such as grasping, insertion, pushing, and dexterous object handling.

See all 413 papers

Facet-0: A Robotic Foundation Model for Contact-Rich Precise Manipulation
Sep 1, 2026Haoyuan Deng, Haichao Liu, Wenkai Guo +6Robot Skill Learning Contact-Rich Robotic Manipulation
1PINE Lab, Nanyang Technological University, Singapore
Proprioception-Anchored Cross-Modal Pretraining for Zero-Shot Sim-to-Real Contact-Rich Assembly
Sep 7, 2026Yuhan Wang, Yurou Chen, Hongye Jiang +2RL for Robotics Cross-Modal Representation Learning
Shanghai Jiao Tong University · University of Sussex
Learning to Build: Autonomous Robotic Assembly of Stable Structures Without Predefined Plans
Feb 27, 2026Jingwen Wang, Johannes Kirschner, Paul Rolland +2Robotic Construction Sequential Decision Making
Lab of Creative Computation, ´Ecole Polytechnique F´ed´erale de Lausanne · Swiss Data Science Center

Sep 1, 2026J/K move · Enter open · S save

Facet-0: A Robotic Foundation Model for Contact-Rich Precise Manipulation

Haoyuan Deng, Haichao Liu, Wenkai Guo, Yuan Ling+5

1PINE Lab, Nanyang Technological University, Singapore

Robot Skill Learning Contact-Rich Robotic Manipulation Robot Foundation Models Robotic RL Force-Controlled Robotic Manipulation Dynamic Robotic Manipulation+1

Open paper AskPDF arXiv

Real-world robotic assembly at sub-millimeter tolerances demands spatial precision, compliant interaction, and robustness to contact failures. We present Facet-0, a robotic foundation model that predicts and values the contact consequences of its actions. Facet-0 unifies multimodal representation learning and reinforcement learning (RL) post-training around a joint action-wrench proposal: a causal wrench history is aligned with vision-language semantics and kinematic state, and flow matching generates each action chunk together with the future wrist-wrench profile it is expected to induce. Deployment rollouts train a distributional Action-Wrench Critic to distinguish motions with similar task progress but different contact outcomes, while phase-aware rewards and contact-selective credit concentrate policy improvement on decisive interactions. To accommodate part-specific dynamics, a lightweight bounded actor reuses the frozen representation for on-robot adaptation; RL remains defined over executable Cartesian actions, while an auxiliary wrench head preserves predictive, non-commanded action-contact coupling. Trained on ManuFacet-1K, a 1,000-hour force-synchronized corpus spanning three embodiments and multiple manufacturing cells, the bounded task-adapted system reaches 82% mean success on five sub-millimeter computer-assembly tasks, compared with 15% for the strongest baseline, with 0.5 mm placement accuracy and 50 ms command latency.

Abstract

Contact-Rich Robotic Manipulation

Explore similar work

Facet-0: A Robotic Foundation Model for Contact-Rich Precise Manipulation

Proprioception-Anchored Cross-Modal Pretraining for Zero-Shot Sim-to-Real Contact-Rich Assembly

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