DAWN: Diffusion-based Alignment and unfastening Workflow Network

Hemant Kumar; Piyush Goenka; Nikhil Chopra

cs.ROOct 29, 2026

DAWN: Diffusion-based Alignment and unfastening Workflow Network

Authors: Hemant Kumar, Piyush Goenka, Nikhil Chopra

Organizations: Department of Mechanical Engineering, University of Maryland, College Park, MD, USA

Abstract

We present DAWN (Diffusion-based Alignment and unfastening Workflow Network), a hierarchical robotic framework designed to perform real-world, unfastening tasks through multi-modal sensory integration. Autonomous robotic unfastening is crucial for improving recycling processes, enhancing safety, and reducing operational costs; yet, it remains predominantly manual due to inherent challenges in precise screwdriver engagement with fasteners, detailed force execution, and robustness to environmental variability. To address the complexity of learning long-horizon, contact-rich manipulation tasks, we subdivide the unfastening task procedure into two sequential stages: (1) alignment & insertion, and (2) unfastening. DAWN consists of three key components: AlignNet, a diffusion-based policy ensuring precise screwdriver tip engagement; LatchNet, a multi-modal classifier reliably detecting successful stage transitions; and UnfastNet, which achieves stable, force-controlled screw removal. Experimental validation on a Franka Emika robot demonstrates that DAWN achieves alignment & insertion success rates comparable to trained human operators, robustly transitions between stages, and attains an impressive success rate in unfastening across various unseen scenarios. Ablation studies further elucidate critical input modalities and effective training strategies, highlighting DAWN's scalability, robustness, and data efficiency.

Figures & tables

Fig. 1: Overview of DAWN, a hierarchical model for solving sequential, contact-rich, long-horizon unfastening tasks.

Fig. 2: The architecture of the hierarchical framework, DAWN, consists of three components: (a) LatchNet, (b) AlignNet, and (c) UnfastNet. (a) LatchNet: Uses ResNet-18 to encode mount images and an MLP encoder for force input to predict stage transitions.(b) AlignNet: Uses ResNet-18 to encode mount images, an MLP encoder for force input, and proprioceptive input. An action token, along with these modalities, is passed through a transformer decoder to obtain an observation representation. This observation representation is then processed by the diffusion policy to generate the action sequence.(c) UnfastNet: Uses an MLP for force input to predict actions.

Input Modalities				AlignNet
Mount Camera	Robot Proprio	Force	Front Camera	Alignment	Insertion	Proximity	Success Rate
Yes	Yes	No	No	$5.0\pm 0.8$	$8.3\pm 1.2$	$8.3\pm 1.2$	$5.0\pm 0.8$
Yes	No	No	No	$1.4\pm 1.2$	$0.0\pm 0.0$	$2.7\pm 2.3$	$0.0\pm 0.0$
Yes	Yes	Yes	No	$0.7\pm 0.4$	$0.7\pm 0.4$	$4.0\pm 0.8$	$0.3\pm 0.4$
Yes	Yes	No	Yes	$1.0\pm 0.0$	$0.0\pm 0.0$	$3.5\pm 0.5$	$0.0\pm 0.0$

TABLE I: Success rates (out of 10) averaged over three seeds for policies trained with different input modalities.

Fig. 3: Common reasons behind unsuccessful policy roll-outs.

	Human Operator			Diffusion Policy	ResNet-18
	Exp 1	Exp 2	Non-Exp 1	W/O Obs-Trunk	Scratch	Finetuned	$4^{th}$ -Layer Finetuned
Alignment	$6.0\pm 0.0$	$6.0\pm 0.0$	$3.0\pm 0.0$	$2.0\pm 0.0$	$5.0\pm 0.8$	$0.0\pm 0.0$	$0.0\pm 0.0$
Insertion	$10.0\pm 0.0$	$10.0\pm 0.0$	$10.0\pm 0.0$	$0.0\pm 0.0$	$8.3\pm 1.2$	$0.0\pm 0.0$	$0.0\pm 0.0$
Proximity	$10.0\pm 0.0$	$10.0\pm 0.0$	$10.0\pm 0.0$	$2.0\pm 0.0$	$8.3\pm 1.2$	$2.0\pm 0.0$	$10.0\pm 0.0$
Success Rate	$6.0\pm 0.0$	$6.0\pm 0.0$	$3.0\pm 0.0$	$0.0\pm 0.0$	$5.0\pm 0.8$	$0.0\pm 0.0$	$0.0\pm 0.0$
Avg. Time (sec)	$23.6\pm 8.2$	$31.75\pm 6.1$	$23.9\pm 5.6$	-	$30.2\pm 4.8$	-	-

TABLE II: We compare the best DAWN policy’s performance against three human teleoperators (two experts and one non-expert) and a baseline diffusion policy [ 30 ] . We also highlight the importance of vision feedback in these precise, contact-rich robotic tasks.

Table 6

Fig. 4: Overhead view illustrating the different target object positions used for training and evaluation. The light green box indicates the extent of positional variation in the training data, and the evaluation locations are labeled. Different types and sizes of fasteners, and materials were employed for assessing DAWN’s performance.

Type	Phillips	Security Torx	Slotted	Hex
Success	5/5	5/5	5/5	5/5

TABLE V: Unfastening success rates across different fastener types.

Robotic Manipulation

Robotic manipulation studies how robots perceive, plan, and control physical interactions to grasp, move, assemble, and use objects.

See all 1,704 papers

CONTACT: CONtact-aware TACTile Learning for Robotic Disassembly
Mar 9, 2026Yosuke Saka, Jyun-Chi Hu, Adeesh Desai +6Robotic Manipulation Contact-Rich Robotic Manipulation
the Department of Industrial Engineering, Purdue University, West Lafayette, IN 47907, USA · the Department of Mechanical Engineering, Texas A&M University, College Station, TX 77843, USA
Force-Driven Hook Disassembly via Constraint-Aware Object-Centric Reinforcement Learning
Oct 28, 2026Nadeen Elhadad, Takuya Kiyokawa, Kensuke HaradaRobotic Manipulation Reinforcement Learning
N. Elhadad, T. Kiyokawa, and K. Harada are with the Department of Systems Innovation, Graduate School of Engineering Science, The University of Osaka, 1-3 Machikaneyama, Toyonaka, Osaka, Japan. · K. Harada is with the Industrial Cyber-physical Systems Research Center, The National Institute of Advanced Industrial Science and Technology (AIST), 2-3-26 Aomi, Koto-ku, Tokyo, Japan.
Learning Forward & Reverse Skills from a Single Unfinished Demonstration for Constrained Manipulation Tasks
Jul 15, 2026Yexin Hu, Haoyi Zheng, Johannes Heidersberger +1Robot Skill Learning Contact-Rich Robotic Manipulation
Autonomous Systems Lab, Technische Universität Wien (TU Wien), Austria · Institute of Robotics and Mechatronics, German Aerospace Center (DLR), Germany

Mar 9, 2026J/K move · Enter open · S save

CONTACT: CONtact-aware TACTile Learning for Robotic Disassembly

Yosuke Saka, Jyun-Chi Hu, Adeesh Desai, Zhiyuan Zhang+5

the Department of Industrial Engineering, Purdue University, West Lafayette, IN 47907, USA · the Department of Mechanical Engineering, Texas A&M University, College Station, TX 77843, USA

Robotic Manipulation Contact-Rich Robotic Manipulation Tactile Sensing for Robotic Manipulation

Open paper AskPDF arXiv

Robotic disassembly involves contact-rich interactions in which successful manipulation depends not only on geometric alignment but also on force-dependent state transitions. While vision-based policies perform well in structured settings, their reliability often degrades in tight-tolerance, contact-dominated, or deformable scenarios. In this work, we systematically investigate the role of tactile sensing in robotic disassembly through both simulation and real-world experiments. We construct five rigid-body disassembly tasks in simulation with increasing geometric constraints and extraction difficulty. We further design five real-world tasks, including three rigid and two deformable scenarios, to evaluate contact-dependent manipulation. Within a unified learning framework, we compare three sensing configurations: Vision Only, Vision + tactile RGB (TacRGB), and Vision + tactile force field (TacFF). Across both simulation and real-world experiments, TacFF-based policies consistently achieve the highest success rates, with particularly notable gains in contact-dependent and deformable settings. Notably, naive fusion of TacRGB and TacFF underperforms either modality alone, indicating that simple concatenation can dilute task-relevant force information. Our results show that tactile sensing plays a critical, task-dependent role in robotic disassembly, with structured force-field representations being particularly effective in contact-dominated scenarios.

Abstract

Robotic Manipulation

Explore similar work

CONTACT: CONtact-aware TACTile Learning for Robotic Disassembly

Force-Driven Hook Disassembly via Constraint-Aware Object-Centric Reinforcement Learning

Learning Forward & Reverse Skills from a Single Unfinished Demonstration for Constrained Manipulation Tasks