OSMO Pipeline: Skill Transfer & Ontology

• OSMO Pipeline is a dual framework that encompasses a tactile glove system for human-to-robot skill transfer and an ontology for semantic workflow management.
• The tactile glove pipeline uses a four-stage process with calibrated 12-sensor arrays and diffusion policy learning to achieve high-precision, contact-rich robot manipulation.
• The OSMO ontology formalizes simulation workflows in computational molecular engineering, enhancing reproducibility, provenance tracking, and automated reasoning.

The term "OSMO pipeline" encompasses two distinct frameworks in contemporary research: (1) the OSMO open-source tactile glove pipeline for human-to-robot skill transfer, and (2) the OSMO Ontology for Simulation, Modelling, and Optimization for semantic interoperability in computational molecular engineering. Both share a focus on structured, reproducible transfers—of skills and of provenance/workflow semantics, respectively—but target different domains and methodologies (Yin et al., 9 Dec 2025, Horsch et al., 2019).

1. OSMO Pipeline for Human-to-Robot Skill Transfer: System Overview

The OSMO pipeline is a four-stage process designed to learn and transfer contact-rich manipulation skills from humans to robots using a shared, sensorized tactile glove system. It eliminates the need for robot-collected data in demonstration, narrows the embodiment gap, and bypasses indirect, vision-only force inference.

Stages:

1. Hardware Design: Development of a wearable glove with twelve three-axis tactile sensors ("taxels")—five on the fingertips, seven across the palm.
2. Data Collection and Signal Processing: Synchronous acquisition of dense tactile signals, RGB/stereo video, and hand kinematics, followed by calibration and augmentation.
3. Policy Learning: Training a diffusion policy model on multimodal observations extracted from human demonstrations.
4. Robot Execution and Transfer: Policy deployment on a robot equipped with the identical glove, including safety-constrained kinodynamic retargeting and force-consistent execution (Yin et al., 9 Dec 2025).

2. Tactile Hardware, Signal Acquisition, and Integration

The OSMO glove features twelve sensor modules, each comprising paired BMM350 3-axis magnetometers mounted beneath a magnetic elastomer. Forces in the range $0.3\,\mathrm{N}$ to $80\,\mathrm{N}$ are detected by mapping raw magnetic flux changes ($\pm2000\,\mathrm{\mu T}$ range, functional band $50$–$1000\,\mathrm{\mu T}$) to normal/shear forces with per-taxel calibration.

Design solutions for electromagnetic crosstalk include differential dual-magnetometer sensing and MuMetal shielding, reducing RMS noise from $\sim150\,\mathrm{\mu T}$ to $\sim35\,\mathrm{\mu T}$ during active manipulation.

The sensor array is polled at $100\,\mathrm{Hz}$ by an STM32 microcontroller, with all streams time-stamped and logged via USB-C. Vision integration leverages low-profile, visually neutral gloves compatible with advanced hand-tracking (e.g., Aria 2, Manus Quantum) and rigid-body coordinate frame transformation to link hand motion with robotic hardware (Yin et al., 9 Dec 2025).

3. Demonstration Logging, Tactile Processing, and Data Preprocessing

Multimodal data (tactile: $12\times3\times2$ flux channels at $100\,\mathrm{Hz}$; video: RealSense D435 RGB/stereo IR at $25\,\mathrm{Hz}$) are acquired synchronously using ROS2, facilitating later alignment for policy learning. Calibration entails baseline subtraction, differential flux calculation, and Savitzky–Golay filtering.

Preprocessing steps include:

• Per-channel percentile normalization and clipping,
• Random center cropping and test-time consistency for image data,
• Histogram balancing of contact intensities to prevent model bias.

All signals are temporally interleaved such that at each keyframe, a 36-dimensional tactile vector $f_t$ is available, aligned with image and pose embeddings (Yin et al., 9 Dec 2025).

4. Policy Learning and Robot Transfer Architecture

Observation vectors $x_t = [f_{1,t}, \dots, f_{12,t}, p_t, z_t]$ concatenate the 36-dimensional tactile force vector, 13-DoF robot pose, and DINOv2 image embeddings. The policy uses a diffusion-based generative model (DDPM)—actions for 16-step robot control chunks are cast as noisy input to be denoised by a U-Net, conditioned via FiLM modulators on the current multimodal observation.

Optimization uses Adam ($\mathrm{lr}=2\times10^{-4}$, batch $=128$, 2000 epochs), with a deployment rate of $2$ Hz (four robot commands per output chunk). Kinematic retargeting from human to robot is achieved by depth-refining human keypoints, solving inverse kinematics with safety constraints for the robot’s joint space.

The identical tactile sensor configuration assures consistency and continuity of the feedback signal, mitigating the embodiment gap without requiring vision-domain inpainting or force estimation (Yin et al., 9 Dec 2025).

5. Experimental Benchmarking and Results

The system was evaluated on a challenging contact-rich task: erasing marker stencils from a $20\,\mathrm{cm}\times20\,\mathrm{cm}$ whiteboard using a sponge-attached Ability Hand. Policy success was measured by pixel-wise marker removal on pre/post-execution image masks.

Sensing Modality	Success Rate (%)
Proprioception only	$27.12 \pm 32.38$
Vision + Proprioception	$55.75 \pm 30.01$
Tactile + Vision + Proprioception	$71.69 \pm 27.43$

Policies using the full OSMO policy (including tactile input) demonstrated improved consistency in maintaining contact and boundary adherence, avoiding under-pressing and drift failure modes observed in vision-only conditions. This suggests that continuous tactile input is critical for contact-dense, high-precision manipulation (Yin et al., 9 Dec 2025).

6. OSMO Ontology Pipeline: Semantic Workflow Interoperability

Independently, the OSMO Ontology for Simulation, Modelling, and Optimization formalizes simulation workflows as machine-interpretable graphs to improve interoperability and provenance in computational molecular engineering (Horsch et al., 2019). It is expressed in OWL-DL, providing:

• Encodings of workflow segments (UseCase, MaterialsModel, Solver, Processor),
• Explicit workflow graphs (ConcreteGraph, VirtualGraph),
• Connective data abstractions (LogicalResource, LogicalVariable, LogicalAccess),
• Properties capturing control-flow, data-flow, concurrency, and causality.

Fine-grained provenance is achieved by mapping logical data accesses (read/write) and resource transfers between workflow nodes. Integration with community ontologies (VISO for software, EMMO for materials models, QUDT for units, IAO/OTRAS for citations) enables semantic linkage and consistency across multi-platform workflows (Horsch et al., 2019).

An example instantiation (phosgene EOS parameterization pipeline) demonstrates mapping simulation, parameter files, and results as individuals in the ontology—enabling automated reasoning, workflow optimization, and provenance-tracing.

7. Community Adoption, Resources, and Interoperability

All OSMO hardware designs, firmware, and codebases supporting the tactile glove pipeline are openly available, with comprehensive documentation for assembly, data acquisition, and policy training (see https://jessicayin.github.io/osmo_tactile_glove/). The OSMO ontology is similarly positioned as a canonical, OWL-DL-compliant intermediate layer for interoperability between workflow management systems (e.g., TaLPas, AiiDA, FireWorks) and data repositories (e.g., VIMMP), facilitating automated schema validation, provenance tracking, and software/method comparison (Yin et al., 9 Dec 2025, Horsch et al., 2019).

A plausible implication is that the OSMO pipeline, in both its tactile robotics and semantic workflow variants, is establishing a rigorous model for reproducible, transferable research practices—both in physical skill transfer and in metadata-driven computational science.