PRISM maps 5,000+ contact-rich industrial robot trajectories

Evidence review - August 2026

PRISM is an August 18, 2026 arXiv preprint and announced multimodal dataset for contact-rich industrial manipulation. The authors report more than 5,000 robot trajectories, an equal number of paired human demonstrations, more than 45 hours of data, and more than 25 tasks across Franka, Realman, and LEJU platforms.

The most important qualification is easy to lose in a headline: tactile sensing is included for only a subset of episodes. The approximately 27 million images reported by the paper combine visual and visuotactile streams; they are not 27 million tactile images. PRISM is therefore a broad industrial robot-learning dataset with a tactile subset, not a uniformly tactile dataset.

Source snapshot

Seven hardware configurations

PRISM combines three robot families, three gripper categories, and three teleoperation interfaces. The paper lists these seven configurations.

The tracker platform uses two Franka Emika Panda arms. The exoskeleton platform uses two Realman RM75-6F arms on a torso with three waist degrees of freedom. The VR platform uses a LEJU upper-body humanoid. These are collection configurations, not evidence that one policy transfers reliably across all three robot families.

Use the Franka Emika Panda robot record for the normalized platform identity. The PRISM paper does not name a commercial tactile-sensor product, so RoboSkin does not invent a sensor relationship for its visuotactile gripper or dexterous hand.

Modalities and native rates

The paper keeps original timestamps because the modalities run at different native rates.

Processed episodes share a common schema containing robot state and action, wrench data when available, multi-view RGB-D, visuotactile imagery when available, calibration parameters, timestamps, platform and task identifiers, outcome labels, and volunteer ratings. The source says experiments convert the data to LeRobot v3.0 format; that experimental conversion should not be treated as proof that the unreleased public package already contains every field in a finalized LeRobot archive.

Collection and evaluation boundary

Eight volunteers collected demonstrations after standardized training, then participated in filtering, annotation, and scoring. The collection also includes intentional perturbation episodes. That adds operator and interaction variation, but it does not by itself establish balanced coverage or leakage-free train and test partitions.

The paper evaluates ACT, Diffusion Policy, and π0 on a bimanual Realman platform for electronic plug/unplug, caliper packaging, and conveyor sorting. It compares 100- and 200-demonstration settings and uses 20 evaluation episodes per configuration. The authors explicitly report that performance remains far from satisfactory, especially for dynamic manipulation and precise force-aware operations.

These experiments are a first-party evaluation of selected PRISM tasks. They are not an independent leaderboard, a universal industrial-robot benchmark, or proof that more demonstrations will improve every task and policy at the same rate.

Availability and license audit

There is a material difference between the paper abstract and the current release surface. The abstract says the dataset is open-sourced at the project page. The official page currently displays a disabled ModelScope dataset button labeled “soon,” while the linked GitHub repository contains the project website, assets, and README but no dataset files, release package, or dataset license.

RoboSkin therefore records PRISM as **announced, download pending** as of August 22, 2026. The license attached to an article or source-code repository must not be assumed to license unreleased dataset files. Availability should change only after an official host exposes files, a version, and reuse terms.

Why PRISM matters for tactile AI

PRISM connects the robotics dataset problem to the tactile AI stack. Vision supplies scene and geometry information, proprioception records robot state, force/torque measures interaction load, and the tactile subset exposes local contact evolution. Its multi-rate timestamps and calibration graph are as important as the headline trajectory count because contact-rich learning depends on aligning these signals without hiding latency or missing modalities.

For robot teleoperation, the three interfaces create a useful research question: how do exoskeleton, tracker, and VR demonstrations differ in precision, smoothness, contact stability, and corrective behavior? The paper motivates that comparison, but the unreleased files prevent an independent answer today.

Release verification checklist

• Confirm the official download host, version, checksum, and dataset-file license.
• Count complete robot trajectories, paired human demonstrations, failed episodes, and per-task coverage from the released manifests.
• Identify exactly which episodes contain tactile imagery, force/torque, depth, and each calibration record.
• Preserve original timestamps and document any resampling or interpolation used for policy training.
• Split complete tasks, episodes, operators, objects, and hardware configurations before extracting overlapping windows.
• Keep results by robot, gripper, teleoperation interface, task, and modality instead of reporting one undifferentiated score.
• Reproduce the ACT, Diffusion Policy, and π0 experiments before comparing PRISM with unrelated datasets.

Related RoboSkin resources

• Tactile robotics datasets
• Robotics datasets
• Robot teleoperation
• Robot learning
• Robot manipulation
• Tactile AI

Primary sources

• arXiv: PRISM - Precision and contact-rich Real-world Industrial Skill dataset with Multimodal sensing
• Official PRISM project page
• Official PRISM GitHub repository