Open Source RoboticsRobots, software & learning

9DTact Compact 3D/6D Vision-Based Tactile Sensor

Robot / hardware build

MIT-licensed compact vision-based tactile sensor providing 3D contact shape reconstruction and 6D force estimation, with SolidWorks/STEP/STL/DXF CAD, LED PCB files, and PyTorch/ROS software. [1] [2]

Build & first use →Official websiteSource codeUpstream build guideBill of materials

For
Robotic manipulation and tactile-sensing researchers and engineers building or integrating compact optical tactile sensors. [1]
Cost
No current cost yet. Upstream bill of materials [1]
Source rights
Named licenses are documented by part. Review licenses
Unknowns
host computer, operating system not fully established. Review gaps

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware
Ecosystem
ROS 1

Requirements · Research details and dates

On this page

Overview

9DTact is a compact vision-based tactile sensor from Changyi Lin et al. (published in IEEE RA-L 2023 / ICRA 2024) designed for both 3D contact shape reconstruction and 6D force estimation. The pinned repository carries a root MIT license and explicitly invites commercial reuse. [1] [2]

Is this for you?

Use cases
  • 3D contact shape reconstruction
  • Generalizable 6D force estimation
  • Simultaneous shape reconstruction and force estimation (SSAF) in ROS

[1]

Known limitations
  • Full mechanical BOM is linked externally as a Chinese Google Doc rather than stored as a priced markdown/CSV BOM in the repository (repository includes only the LED PCB BOM spreadsheet).
  • Training a custom 6D force estimator requires collecting paired wrench data with an external F/T reference sensor (e.g., BOTA MiniONE Pro).
  • LED PCB editable source is in EasyEDA Pro (.epro) format rather than KiCad, and core mechanical CAD assemblies are SolidWorks (.SLDPRT/.SLDASM) alongside STEP/STL/DXF fabrication exports.
  • Source review only: elastomer molding, camera calibration, and force-estimation accuracy were not physically tested.

Scoped to inspected repository files and README.

[1] [2] · Editorial assessment

What you need

Dependencies
  • Python 3.8
  • PyTorch (CUDA recommended for training)
  • ROS (optional, for shape-force_ros nodes)
  • BOTA MiniONE Pro driver (optional, for new 6D force dataset collection)

[1]

Tools
  • 3D printer (sensor shell, base, isolation ring, calibration board)
  • Laser cutter / CNC for 2.5–3.0 mm DXF gel molds
  • Silicone elastomer casting supplies
  • PCB fabrication service for LED board

[1]

Hardware & parts

Upstream bill of materials [1]

Hardware configuration

Compact optical fingertip tactile sensor module

[1]

Sensors
  • Internal compact USB camera module (frank-s15-v1.0 referenced in CAD)
  • Translucent/black elastomer gel layer with custom LED illumination ring

[1]

Assembly & first use

Open the upstream getting-started guide [1] · Project documentation [1]

  • Fabricate the sensor base, shell, isolation ring, and gel molds from 9DTact_Design/fabrication/ and LED illumination PCB from 9DTact_Design/LED_PCB/.
  • Create a Python 3.8 conda environment with PyTorch and run pip install -e . in the repository.
  • Adjust camera lens focus (~15 mm), print calibration_board.STL, and run _1_Camera_Calibration.py followed by _2_Sensor_Calibration.py using a ~4.0 mm radius ball.
  • Run _3_Shape_Reconstruction.py, _1_Force_Estimation.py, or the ROS nodes in shape-force_ros/.

[1]

Assembly

Three-module design in 9DTact_Design/: Perception Module (camera + LED PCB + sensor base), Diffusion Module (acrylic window + isolation ring + shell + transparent gel), and Contact Module (translucent + black gel layers).

[1]

Calibration

Camera calibration via printed calibration_board.STL and sensor depth calibration via a ~4.0 mm radius calibration ball (not required if using the pre-calibrated production unit).

[1]

Cost

No current, complete configuration-specific build cost has been established for this profile.

Cost and review boundaries

No total DIY hardware cost is established in the pinned repository: the full sensor BOM is linked externally as a Chinese Google Doc, and an external Taobao production unit link is also referenced. This review inspected only the pinned GitHub revision (453bd85094ef); no parts were fabricated, no silicone gels were cast, and neither the calibration pipeline nor the ROS/PyTorch models were executed on hardware. [1] [2]

3D preview

Preview not available yet. The upstream design is linked below. An interactive preview has not been prepared for this profile.

Project notes

Mechanical, PCB, and software artifacts

The 9DTact_Design/ tree splits the hardware into three subassemblies—Perception Module (camera module frank-s15-v1.0, LED board, and sensor base), Diffusion Module (acrylic window, isolation ring, shell, and transparent gel), and Contact Module (translucent and black gel layers). Alongside SolidWorks .SLDPRT/.SLDASM files, 9DTact_Design/fabrication/ supplies STEP and STL exports for the housing, isolation ring, and calibration board, plus DXF templates for 2.5 mm, 2.8 mm, and 3.0 mm gel-casting molds. 9DTact_Design/LED_PCB/ adds the EasyEDA Pro project (.epro), Gerber archive, pick-and-place file, and LED PCB BOM. [1]

The documented DIY workflow moves from part fabrication and multi-layer gel casting to setting up the Python 3.8 / PyTorch environment, focusing the camera at roughly 15 mm, running lens-focus camera calibration (calibration_board.STL), performing sensor depth calibration with a recommended 4.0 mm radius ball (adjusted for surface thickness), and then running 3D shape reconstruction. Upstream notes that the commercial production version skips both camera and sensor calibration. The repository also includes DenseNet-169 6D force estimation (with a saved checkpoint in saved_models/) and ROS nodes for simultaneous shape and force estimation, while retraining force estimators on new hardware requires paired ground-truth wrench collection using an external force/torque sensor such as a BOTA MiniONE Pro. [1]

What we checked

Source review only. No physical sensor fabrication, gel casting, calibration, or PyTorch/ROS runtime test was performed. No local image asset is bundled yet.

Not yet established

  • Requirements: operating system, host computer

Updates

Upstream last updated
Aug 21, 2026 [1]
Latest release
Not established
Profile updated
Nov 21, 2026
Sources last checked
Nov 21, 2026
Next review
Dec 21, 2026
Reviewed revision
453bd85 · combined hardware and software repository

Source files and licenses

Permissively MIT-licensed tactile sensor with complete mechanical CAD (STEP/STL/DXF/SolidWorks), LED PCB fabrication files, PyTorch/ROS software, and a pretrained DenseNet-169 checkpoint; partial because the camera module is bought-in, full sensor BOM is external, and raw training datasets are not bundled.

LayerAvailability / license
mechanical designavailable · MIT
SolidWorks assemblies/parts plus STEP, STL, and DXF fabrication/mold files under MIT. [1] [2]
electronicsavailable · MIT
LED illumination board EasyEDA Pro source (.epro), Gerbers, BOM, and pick-and-place files; USB camera module is bought-in. [1] [2]
softwareavailable · MIT
Python/PyTorch calibration, 3D shape reconstruction, 6D force estimation, and ROS integration code. [1] [2]
model weightsavailable · MIT
Saved DenseNet-169 PyTorch checkpoint included in saved_models/. [1] [2]
datasetsnot found · License not fully assessed
Dataset loader and collection scripts are included, but raw tactile-force training datasets are not stored in the repository. [1]
documentationavailable · MIT
README, linked arXiv paper (2308.14277), and YouTube video tutorial. [1] [2]

Uses a standard USB camera module and passive LED illumination board with host-side processing; no custom microcontroller firmware.

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.7.0-draft.5

Change history

2026-11-21 — Initial source-reviewed draft for Hardware Blind-Spot Batch 1.