Code and data for estimating the roll angle of the human forearm (pronosupination, denoted q5) from the proprioceptive state of an underactuated robotic gripper while it grasps the forearm — i.e. from the joint angles of the gripper's passive phalanges, without vision.
This repository supports the tactile-proprioception part of the Master's thesis:
Sistema multisensorial para la interacción física humano-robot Rodrigo Castro Ochoa. Tutor: Jesús Manuel Gómez de Gabriel. Máster en Ingeniería Mecatrónica, Escuela de Ingenierías Industriales, Dpto. de Ingeniería de Sistemas y Automática, Universidad de Málaga, 2026.
The method is described in Section 2.3 (theoretical framework) and
Section 3.4 (method) of the thesis; the datasets here come from
Chapter 4, Experiments 2 and 3. A transcription of Section 2.3 is in
docs/section_2.3.md.
Strictly visual pose estimation cannot observe forearm pronosupination: the detected skeleton ends at the wrist. When the robot's gripper grasps the forearm, the cross-section of the forearm can be modelled as an ellipse (Ref. [4] in the thesis: 3D scans of 6 subjects, R² 0.898–0.980, RMSE 0.21–0.64 % of the forearm circumference). The mechanical envelope formed by the gripper's under-actuated links is a convex contact polygon; fitting an ellipse inside that polygon recovers the forearm section. Doing this for the two parallel pinches of the gripper (separated 75 mm) gives two ellipse centres, hence the forearm longitudinal axis (the q5 rotation axis) and, from the orientation of the major diagonal of the distal ellipse, the angle q5.
Thesis fig. 3.5 — from the passive phalange angles (θ₁, θ₂) the gripper forms a
convex contact polygon; the inscribed ellipse (centre G) models the forearm
section. More figures in docs/method.md.
Two ellipse-fitting strategies are implemented:
| Method | Idea | Notes |
|---|---|---|
| John ellipse (MVIE) | Maximum-area ellipse inscribed in the contact polygon, as a convex program max log det(G) s.t. G ⪰ 0, ‖G aᵢ‖₂ + aᵢᵀc ≤ bᵢ. |
Fast, global optimum; accuracy depends on how well the gripper accommodates the arm. Tends to over-round (low aspect ratio). |
| Fit Anatomical | Fit an ellipse of known anthropometric aspect ratio a/b; optimise centre (cx,cy), orientation θ (= q5) and isotropic scale s with an asymmetric penalty, solved with L-BFGS-B. |
Robust to sub-optimal grasps. Quantitative validation is reported in the forthcoming journal paper (see CITATION.cff). |
.
├── config/
│ └── gripper.yaml # link lengths, pinch separation, erosion distance, solver bounds
├── code/
│ ├── acquisition/ # ROS / rosserial logging of phalange encoders + IMU ground truth
│ ├── estimation/ # the proprioceptive pipeline (importable Python package)
│ │ ├── gripper.py # forward kinematics + contact polygon (hexagon/pentagon/rhombus)
│ │ ├── erosion.py # morphological erosion (Minkowski difference, Shapely buffer, d = -7.5 mm)
│ │ ├── mvie.py # maximum-area inscribed ellipse (convex optimisation, "John ellipse")
│ │ ├── fit_anatomical.py # asymmetric-penalty anatomical fit (L-BFGS-B)
│ │ ├── forearm_axis.py # two ellipse centres -> forearm axis, elbow & wrist localisation
│ │ └── pronosupination.py # q5 from the distal ellipse major diagonal and the {arm} frame
│ ├── processing/ # raw logs -> tidy per-trial tables
│ └── analysis/ # error analysis and figures for Experiments 2 and 3
├── data/
│ ├── raw/ # NOT stored in this repo — see "Data" below
│ └── processed/ # regenerable tidy tables (not tracked)
├── docs/
│ ├── section_2.3.md # transcription of the thesis section
│ ├── method.md # full derivation: polygon -> erosion -> ellipse -> q5
│ └── protocol.md # Experiment 2 & 3 protocols
└── results/
├── figures/
└── tables/
pHRI workstation (thesis §2.1): Franka FR3 collaborative manipulator with an
under-actuated gripper of two parallel pinches designed to grasp a
participant's forearm; 4 extrinsically calibrated RGB-D cameras (eye-to-hand);
ROS Noetic. Each gripper finger: proximal phalanx L1 = 40 mm, distal phalanx
L2 = 50 mm; the two pinches are 75 mm apart; phalange angles read by
high-resolution magnetic encoders. Full gripper kinematics/dimensions:
Ruiz-Ruiz et al., Compliant gripper with force estimation for physical
human–robot interaction, Mechanism and Machine Theory, 2022, 178, 105062
(https://doi.org/10.1016/j.mechmachtheory.2022.105062). See
hardware/README.md.
Ground truth for q5: an IMU / accelerometer (MPU module) held by the
participant, giving the roll angle relative to the end-effector frame
(q5_GT = atan2(a_y, a_z) − offset − π/2), logged over rosserial and
synchronised in ROS. Sign convention: positive = supination, negative =
pronation, zero at the neutral pose.
The raw per-participant recordings for Experiments 2 and 3 are not stored in
this repository. They are archived as a separate dataset on Zenodo under
CC-BY-4.0: 10.5281/zenodo.22551414
(see data/README.md for the file schema). Participants
are anonymised (P01, P02, …); no identifying data is stored. Data
collected under signed informed consent with guaranteed anonymisation
(thesis §4.2).
- Experiment 2 (discrete): cohort of 9 (5 M, 4 F). From neutral (0°), forearm resting on the pinch base, static sweep in 20° increments to maximum pronation and then to maximum supination; the gripper is fully opened and closed between every capture to reduce soft-tissue hysteresis. Also records contact-polygon feasibility (hexagon 100 %; pentagon/rhombus ≈ 80 %).
- Experiment 3 (continuous): continuous pronation→supination sweep with the
gripper closed throughout; tactile q5 filtered with a 1-D constant-velocity
Kalman filter (state
[q5, q5_dot],Q = 0.05); pentagon + Fit Anatomical with the ellipse semi-axes calibrated from the participant's real anthropometry.
python -m venv .venv
# Windows: .venv\Scripts\activate | Linux/macOS: source .venv/bin/activate
pip install -r requirements.txt
# Download the dataset from Zenodo (see data/README.md) and unpack it under
# data/raw/ before running the pipeline below.
python code/processing/build_dataset.py # data/raw -> data/processed
python code/analysis/exp2_discrete.py # -> results/figures, results/tables
python code/analysis/exp3_continuous.pySee CITATION.cff.
This work is part of project PID2021-127221OB-I00 (CONCERTO — Control Colaborativo para Interacción física Empática entre RoboT y humanO), funded by MICIU/AEI/10.13039/501100011033/FEDER, UE. TaISLab, Universidad de Málaga.
Este proyecto es parte del proyecto PID2021-127221OB-I00, financiado por MICIU/AEI/10.13039/501100011033/FEDER, UE.
- Code: MIT — see
LICENSE. - Data: Creative Commons Attribution 4.0 International (CC-BY-4.0) — see
data/LICENSE.md.
Rodrigo Castro Ochoa <rcastro@uma.es>, Jesús Manuel Gómez de Gabriel <jesus.gomez@uma.es>, Cristina Urdiales, Óscar de Cózar, Beatriz Blázquez — TaISLab, Universidad de Málaga.
Contact: Rodrigo Castro Ochoa <rcastro@uma.es> · Jesús Manuel Gómez de Gabriel <jesus.gomez@uma.es>.