LABIT: Long-Horizon Robotic Assembly Benchmark for Industrial Tasks

Constantin Schempp; Christian Friedrich; Luca Kofler; Laura Suhm

cs.ROOct 12, 2026

LABIT: Long-Horizon Robotic Assembly Benchmark for Industrial Tasks

Authors: Constantin Schempp, Christian Friedrich, Luca Kofler, Laura Suhm

Organizations: Institute for Robotics and Intelligent Production Systems, University of Applied Sciences Karlsruhe (HKA), Moltkestraße 30, Karlsruhe, 76133, Germany

Abstract

Robot learning has undergone significant advancements in the domain of compliant control and robotic assembly. The development of novel methodologies has transcended the capacity to address discrete tasks, such as grasping or inserting, and has evolved to encompass the comprehensive construction of entire assemblies. Conventional benchmarks, which are predominantly focused on single tasks, are no longer adequate to address the complexity and interconnected nature of modern manufacturing processes. Consequently, the necessity for benchmarks to encompass a broader spectrum of assembly and subassembly steps is becoming apparent. In light of these developments, we propose LABIT, an industrial robotic assembly benchmark for long-horizon manipulation. This benchmark encompasses DIN-norm assembly tasks, including screwing, placing, inserting, and snapping, under both high and low tolerance conditions. The successive nature of these assembly steps implies that prior errors can influence subsequent manipulation. Consequently, this benchmark is particularly relevant for assessing the robustness of intelligent robotic systems. The proposed benchmarking protocols and novel performance metric, based on a taxonomy of assembly complexity in conjunction with the comprehensive list of parts and their corresponding CAD models, serve as essential tools for research and development in the field. To provide a baseline, we evaluate the benchmark using traditional force-position control and report reliability and score metrics per primitive, subtask, assembly category, and subassembly. Accompanying data is available at https://irp-hka.github.io/robotic-assembly-benchmark/.

Figures & tables

Figure 1: Experimental setup for LABIT. A robot arm equipped with a parallel-jaw gripper, a wrist-mounted RGB-D camera, and a force-torque sensor executes the assembly tasks on a dedicated worktable. The benchmark housing is secured in a fixture; individual parts are provisioned via a part feeder, corresponding to planning complexity level 1.

Panel of Figure 2 — Figure 2: Hierarchical structure of the LABIT benchmark. (a) shows explosion views (top row) and corresponding primitive-level assembly graphs (bottom row) for three exemplary assemblies: cover plate (left), PCB module (center), and middle housing (right). Nodes represent individual assembly parts, where node colors match with part colors in the explosion views; edges display the spatial relationship and type of connections. The SCREW_M5x16 nodes of the housing middle assembly are not depicted here, because they are part of a later assembly stage. (b) shows the top-level graph connecting all assembly parts from LABIT, where the highlighted areas correspond to the assemblies shown in (a).

Benchmark	Ops.	Seq.	Ind.	Force	Reor.	Sim.	Phys.
LABIT (ours)	7	●	◐	●	●	●	●
FurnitureBench [ 16 ]	4	●	○	○	◐	●	●
FMB [ 17 ]	3	◐	○	◐	◐	●	●
RAMP [ 18 ]	3	●	◐	○	○	●	●
REASSEMBLE [ 19 ]	4	○	●	●	○	○	●
AutoMate [ 20 ]	2	○	○	○	○	●	●

Table 1: Comparison of robotic assembly benchmarks. Ops. : number of distinct operation types. Seq. : long-horizon sequential assembly with hierarchical step dependencies (failure in step

i

directly affects step

i{+}1

). Ind. : industrial off-the-shelf parts used. Force : force-sensitive/compliant control required. Reor. : explicit subassembly reorientation required. Sim. : simulation environment provided. Phys. : physical hardware provided. ● yes ◐ partial ○ no.

Panel of Figure 3 — Figure 3: Different initial benchmark setup depending on the chosen complexity level. (a) level 1, parts are provided by a part feeder, (b) level 2, parts are provided unoriented on a plate and (c) level 3, parts are provided cluttered in a box.

Panel of Figure 4 — Figure 4: Overview of assembly tasks required to successfully assemble the benchmark. Tasks occurring multiple times like peg insertion are only depicted once. Blue colored objects show the interacting components during task execution.

Figure 5: Taxonomy for the categorical assessment of the complexity of robot-guided assembly tasks. The complexity of each assembly task can be evaluated using the three properties tolerance

T

, geometry

\Omega

and material combination

M

, where

p_{i}

encodes the complexity of each group. In the latter, the colors stand for the stiffness of the material, namely for soft, for medium and for hard.

Figure 6: Examples of point allocation for different geometric complexities

\text{ord}(V)

. The blue line indicates the orientation of the object.

$E_{\mathrm{eff}}$ in GPa	$p_{M}$	Regime	Typical pairings
$<0.1$	3	Large deformation	Elastomer involved (o-ring, tube)
$0.1$ – $1$	2	Noticeable deformation	Soft thermoplastics (PE, PTFE)
$1$ – $10$	1	Rigid, force-tolerant	Engineering polymers, polymer–metal
$10$ – $100$	2	Increased force sensitivity	Light metals (Al–Al, Al–steel)
$\geq 100$	3	High force sensitivity	Steel–steel, ceramics

Table 2: Point allocation for the material combination based on the effective Young’s modulus

E_{\mathrm{eff}}

according to ( 8 ).

Figure 7: Young’s modulus over density for families of engineering materials (after [ 37 ] ) with the bin edges of the material combination. Solid lines mark the primary edges placed in the gaps between the elastomer, polymer and metal families; dashed lines mark the secondary edges obtained by decade subdivision. The resulting points

p_{M}

are given on the right axis.

Symbol	Meaning
$p_{ik}$	complexity points of property $k$ of primitive $i$
$P_{i}$	total points of primitive $i$
$R_{i}$	reliability of primitive $i$
$R_{task}$	reliability of subtask ( $=\prod_{i\in task}R_{i}$ )
$R_{sa}$	reliability of subassembly
$R_{c}$	reliability of primitive category

Table 3: Overview of the metrics used in this work. The table defines symbols for complexity points (

p_{ik}

), reliability measures (

R_{i}

R_{sa}

R_{c}

R

), scores (

S_{i}

S_{sa}

S_{c}

S

), timing metrics (

T_{sa}

T

), and force interaction metrics (

E_{xy}

E_{z}

\sigma_{xy}

\sigma_{z}

Figure 8: Benchmark setup in real environment (left) and simulation (right), showing the robot, gripper, wrist camera and force-torque sensor. The parts are provided by the part feeder, corresponding to planning complexity level 1.

Figure 9: Convex decomposition of assembly parts for accurate contact calculation in simulation. Areas where contact is expected during assembly are modeled with higher detail by using smaller sphere radii (right).

Figure 10: Visualization of executed primitive sequences as seen by the wrist-mounted camera during baseline execution. Each row shows consecutive frames spanning the approach, contact, and completion phase of a single primitive. Top row: PCB insertion, where the board is grasped at its connector socket, aligned above the spring-loaded pins, and pushed into its seat. Center row: o-ring placement, where the ring is picked from the provisioning board, transported to the housing notch, and seated around the rotor shaft. Bottom row: screwing, where the screw is grasped, aligned with the through hole, and driven in by rotating the end effector about the assembly axis.

Parameter	Value
Timestep	0.5 ms (2 kHz)
Integrator	implicitfast
Solver	Newton
Solver tolerance	$10^{-10}$
Gravity	$[0,\,0,\,-9.81]$ m/s 2
Friction cone	elliptic

Table 4: MuJoCo simulation parameters used in LABIT.

Panel of Figure 11 — Figure 11: Results of the baseline implementation using the task-level metrics from Section 3.3.2 . Top: Scores achieved for each primitive, assembly category and subassembly. Bottom: Reliability for each primitive, assembly category and subassembly.

Figure 12: Force interaction metrics for each subtask executed by the baseline method in the real environment. The first and second rows show the force energy in

z

-direction and

xy

-direction and the third and fourth rows show the force smoothness in

z

-direction and

xy

-direction respectively. The letter in parentheses after the component name refers to the visual representation of the subtask in Fig. 4 .

Panel of Figure 13 — Figure 13: Example failure cases during execution of the baseline. (a) inserting a positioning pin fails due to jamming, (b) misalignment of the middle housing during placement, (c) insertion of the internal plug fails due to misalignment, (d) placement of the top housing fails due to slipping of the housing in the gripper, (e) screwing fails due to misalignment of the screw with the through hole and (f) reorienting and placing the module fails due to slipping of the housing in the gripper.

Panel of Figure 14 — Figure 14: Comparison of the forces incurred during task execution in the real environment (top row) and the simulated environment (middle row). The bottom row shows the corresponding force interaction metrics. (a) insertion of gear wheel 1; (b) alignment and insertion of the top housing onto the middle housing; (c) insertion of the outer plug into the socket on the middle housing; and (d) screwing together of the top and middle housings.

Robotic Assembly

Robotic assembly studies how robots perceive, plan, and join parts reliably, from precision insertion to collaborative construction.

See all 87 papers

Jun 2, 2026

WorkBenchMark: A LEGO-Based Assembly Benchmark with an Assembly-by-Disassembly Baseline for the Smart Manufacturing League

We introduceWorkBenchMark, a LEGO Duplo-based robotic assembly benchmark motivated by the RoboCup Smart Manufacturing League. Robotic assembly couples low-level manipulation with task-level symbolic reasoning under physical constraints, a combination that current end-to-end learning methods do not yet solve reliably. The benchmark provides 400 tasks across four complexity tiers. We provide an open-vocabulary perception, Assembly-by-Disassembly baseline solution. Our planning-based pipeline outperforms a modern vision-language-action approach across all tiers. The benchmark, simulation environment, and baseline implementation will be released openly to support the broader robotic assembly community.

Wenbo Ma, Daniel Swoboda, Matteo Tschesche +1

Chair of Machine Learning and Reasoning (i6), RWTH Aachen University, Aachen, Germany · MASCOR Institute, FH Aachen University of Applied Sciences, Aachen, Germany

Robotic Manipulation Robot Task Planning Robot Manipulation Benchmarks Robotic Assembly

Ask PDF

Mar 16, 2026

RoCo Challenge at AAAI 2026: Benchmarking Robotic Collaborative Manipulation for Assembly Towards Industrial Automation

Embodied Artificial Intelligence (EAI) is rapidly developing, gradually subverting previous autonomous systems' paradigms from isolated perception to integrated, continuous action. This transition is highly significant for industrial robotic manipulation, promising to free human workers from repetitive, dangerous daily labor. To benchmark and advance this capability, we introduce the Robotic Collaborative Assembly Assistance (RoCo) Challenge with a dataset towards simulation and real-world assembly manipulation. Set against the backdrop of human-centered manufacturing, this challenge focuses on a high-precision planetary gearbox assembly task, a demanding yet highly representative operation in modern industry. Built upon a self-developed data collection, training, and evaluation system in Isaac Sim, and utilizing a dual-arm robot for real-world deployment, the challenge operates in two phases. The Simulation Round defines fine-grained task phases for step-wise scoring to handle the long-horizon nature of the assembly. The Real-World Round mirrors this evaluation with physical gearbox components and high-quality teleoperated datasets. The core tasks require assembling an epicyclic gearbox from scratch, including mounting three planet gears, a sun gear, and a ring gear. Attracting over 60 teams and 170+ participants from more than 10 countries, the challenge yielded highly effective solutions, most notably ARC-VLA and RoboCola. Results demonstrate that a dual-model framework for long-horizon multi-task learning is highly effective, and the strategic utilization of recovery-from-failure curriculum data is a critical insight for successful deployment. This report outlines the competition setup, evaluation approach, key findings, and future directions for industrial EAI. Our dataset, CAD files, code, and evaluation results can be found at: https://rocochallenge.github.io/RoCo2026/.

Haichao Liu, Yuheng Zhou, Zhenyu Wu +14

Nanyang Technological University, Singapore · Beijing University of Posts and Telecommunications, Beijing, China · Carnegie Mellon University, Pittsburgh, USA +1

Robot Manipulation Benchmarks Robotic Assembly Industrial Robotics Long-Horizon Robotic Manipulation Multi-Task Robotic Manipulation

Ask PDF

Oct 12, 2026

LIBERO-Industry: Benchmarking Robotic Manipulation under Industrial Process Constraints

Industrial manipulation in flexible manufacturing, e.g., high-mix, low-volume (HMLV) electronics assembly, requires robot policies to adapt to changing products and production conditions. Such settings impose industrial process constraints, including product variants, moving material flows, finite execution windows, and tight geometric tolerances. However, existing evaluation frameworks do not systematically capture these constraints when assessing generalist robot policies. To address this gap, we introduce LIBERO-Industry, a novel benchmark built on the widely used robot learning benchmark LIBERO, that translates the aforementioned industrial constraints into controllable task conditions and measurable success criteria. Specifically, LIBERO-Industry comprises three classic task scenarios: Static-to-Dynamic (S2D) loading, Dynamic-to-Static (D2S) unloading, and Dynamic-to-Dynamic (D2D) transfer. The suites contain 216 tasks in total that systematically vary object geometry, conveyor speed, interaction-window duration, and geometric acceptance region size. Extensive experiments across five generalist robot policies (including vision-language-action models and world-action models) show that performance progressively degrades as industrial constraints are introduced and combined. Under a shared fine-tuning and evaluation protocol, the LIBERO-Industry co-training condition achieved higher observed target-transfer completion than the original-LIBERO condition.

Ming Shin Cindy Sung, Duan Leijiahao, Haoxiang Wang +9

Institute of AI for Engineering, Tongji University · The Chinese University of Hong Kong · University of Aberdeen +1

Robotic Manipulation Robot Manipulation Benchmarks Industrial Robotics Generalization in Robotic Manipulation Dynamic Robotic Manipulation

Ask PDF

Abstract

Robotic Assembly

Explore similar work

WorkBenchMark: A LEGO-Based Assembly Benchmark with an Assembly-by-Disassembly Baseline for the Smart Manufacturing League

RoCo Challenge at AAAI 2026: Benchmarking Robotic Collaborative Manipulation for Assembly Towards Industrial Automation

LIBERO-Industry: Benchmarking Robotic Manipulation under Industrial Process Constraints