FIRM: A Benchmark for Industrial Flexible-Object Robot Manipulation

FIRM is a benchmark for industrial flexible-object robot manipulation grounded in real-world industrial data. The benchmark focuses on manipulation tasks involving mixed-stiffness objects, including instruction manuals, power cables, sponge pads, tapes, and cardboard components. These objects exhibit bending, slipping, rolling, compression, elastic recovery, and flexible-rigid contact under production-line constraints.

This repository provides a reviewer-accessible release of the FIRM benchmark dataset. The current release includes representative folders from real industrial teleoperation demonstrations and task-specific manipulation scenarios. The dataset is intended for policy training, offline evaluation, diagnostic analysis, and deformation-aware assessment of industrial manipulation policies.

Dataset Summary

FIRM contains expert teleoperation demonstrations collected from active production-line settings and a high-fidelity physical replica. The demonstrations are recorded with multi-view RGB-D observations, robot state, gripper state, and action sequences. The benchmark supports evaluation beyond binary success using the Deformation-aware Assessment Protocol (DAP), which includes:

• Completion quality
• Deformation-aware execution quality
• Perturbation robustness
• Failure-mode diagnosis

Folder Structure

The uploaded release is organized by task/source folders:

FIRM/
├── Cable&Mouse/
├── Manual_Perturbation/
├── Manual1/
├── Manual2/
├── Manufacturing_Lines_Box/
├── Manufacturing_Lines_Cable&Mouse/
├── Manufacturing_Lines_Manual/
├── Sponge_Pad/
└── Tape/

Folder Descriptions

Folder	Description
Cable&Mouse/	Demonstrations involving cable and mouse-related manipulation, including flexible cable routing and placement.
Manual_Perturbation/	Instruction manual manipulation episodes with pose or scene variations for robustness evaluation.
Manual1/	Instruction manual insertion demonstrations collected in the physical replica setting.
Manual2/	Additional instruction manual insertion demonstrations collected in the physical replica setting.
Manufacturing_Lines_Box/	Real production-line demonstrations involving box or cardboard manipulation.
Manufacturing_Lines_Cable&Mouse/	Real production-line demonstrations involving cable and mouse-related manipulation.
Manufacturing_Lines_Manual/	Real production-line demonstrations involving instruction manual insertion.
Sponge_Pad/	Sponge pad placement demonstrations involving deformable volumetric objects.
Tape/	Tape manipulation demonstrations involving closed-loop objects with rolling and slipping behavior.

Task Categories

FIRM covers five representative industrial flexible-object manipulation categories:

Task Category	Object Type	Core Challenge
Instruction Manual Insertion	Planar sheet	Alignment, confined insertion, folding/creasing avoidance
Cable Manipulation	1D flexible object	Bending stiffness, routing, tangling avoidance
Box Folding	Articulated structure	Pressing, folding, hinge-like deformation
Sponge Pad Placement	Volumetric deformable object	Compression, recovery, placement coverage
Tape Manipulation	Closed-loop object	Rolling, slipping, orientation control

Data Modalities

Depending on the task folder and recording setup, each episode may include:

• Multi-view RGB-D observations
• End-effector poses
• Gripper states
• Delta end-effector actions
• Frame indices
• Task metadata
• Success or completion annotations when available

The dataset is designed to be compatible with imitation-learning and vision-language-action policy pipelines. Processed trajectories may be serialized into LeRobot-style formats in downstream use.

Evaluation Protocol

FIRM is evaluated using DAP, a deformation-aware assessment protocol. DAP reports binary success as a conventional reference, but emphasizes more diagnostic metrics.

Completion Quality

Completion quality measures how much of the task is completed before termination. Examples include:

• Whether an instruction manual is fully inserted, aligned, and not visibly skewed
• Effective cable length contained in the target region
• Target-region coverage for sponge pad placement
• Final position and orientation correctness for tape manipulation
• Folding completion induced by the pressing tool for box folding

Deformation-aware Execution Quality

DAP does not penalize deformation itself. Some deformation is normal or necessary: sponge pads may compress during grasping, cables and tapes may bend under gripper contact, and cardboard components must be folded to complete the task.

Instead, this metric focuses on out-of-tolerance physical outcomes, such as:

• Manual remaining folded after insertion
• Manual corner being pinched or trapped
• Sponge pad remaining folded or compressed after placement
• Sponge corner being trapped under another object
• Residual cable tangling
• Uncontrolled tape rolling
• Rebound-induced displacement
• Jamming, slipping, or dropping

Perturbation Robustness

Perturbation robustness evaluates whether a policy remains effective under controlled changes in object pose, receptacle pose, fixture position, material response, or scene configuration.

Failure-mode Diagnosis

Failure-mode diagnosis categorizes failed or partially completed episodes into interpretable failure modes, such as:

• Grasp failure
• Alignment failure
• Incomplete insertion or placement
• Out-of-tolerance deformation
• Slip or rolling
• Contact instability
• Collision or jamming
• Over-compression
• Timeout

Intended Use

This dataset is intended for:

• Industrial robot manipulation benchmarking
• Imitation-learning policy training
• Vision-language-action policy adaptation
• Offline policy evaluation
• Data-scaling analysis
• Failure-mode diagnosis
• Deformation-aware manipulation assessment

Out-of-Scope Use

This dataset is not intended for:

• Human identification or surveillance
• Safety-critical deployment without additional validation
• Inferring private information about operators or manufacturing sites
• Training policies for deployment without task-specific safety testing
• Reconstructing proprietary production processes beyond the released benchmark tasks

License

The FIRM-Real dataset is released under the Creative Commons Attribution 4.0 International License (CC-BY-4.0), unless otherwise specified in individual files or metadata.

If any curated third-party segments or derived metadata are included in future releases, their redistribution terms will follow the corresponding source licenses. When redistribution of original third-party assets is not permitted, only metadata, annotations, or links will be released.

Responsible AI and Ethics

The dataset contains robot observations, robot states, and robot action trajectories collected in controlled industrial manipulation settings. It does not intentionally contain personally identifiable information. Human operators participate only as teleoperators, and the released data are intended to focus on robot-object interaction rather than human identity.

Potential positive impacts include improving robot manipulation in hazardous, repetitive, or ergonomically challenging industrial tasks. Potential risks include misuse in unsafe automation deployment or labor displacement without appropriate oversight. We encourage responsible use, task-specific safety validation, and consultation with domain experts before real-world deployment.

Citation

A citation will be added after paper review. During anonymous review, please cite this dataset as:

@misc{firm2026anonymous,
  title        = {FIRM: A Benchmark for Industrial Flexible-Object Robot Manipulation},
  author       = {Anonymous Authors},
  year         = {2026},
  howpublished = {Anonymous dataset release for review}
}