DIJIT: A Robotic Head for an Active Observer

Technical Summary: DIJIT: A Robotic Head for an Active Observer

Problem Statement
Robotic vision systems typically differ significantly from human visual systems in terms of degrees of freedom (DOF) and task versatility. While human vision employs active control with high DOF (three rotational DOF per eye plus a controllable neck) to solve diverse visual tasks, most robotic systems utilize fewer DOF, often relying on fixed stereo pairs or simple pan-tilt mechanisms. Existing biomimetic robotic heads frequently lack the necessary kinematic structure (specifically independent pan, tilt, and roll for each camera), possess inter-camera baselines that are too large for human-like performance, or rely on heavy metal frames that hinder mobility. Furthermore, many existing designs do not detail their performance metrics, particularly regarding saccadic eye movements, which are critical for biological active vision. There is a need for a robotic platform that mirrors the physical properties of the human visual system to explore the differences between biological and computer vision methods.

Methodology and System Design
The authors present DIJIT, a novel binocular robotic head designed to replicate human-like eye and head-neck motions.

• Mechanical Design: DIJIT features nine mechanical degrees of freedom: three for the neck (pan, side-bending, flexion/extension) and three for each camera (pan, tilt, and roll). The design achieves a human-like inter-camera baseline of 115 mm. The head is constructed using lightweight PA6-CF 3D-printed parts, weighing approximately 1.4 kg, with dimensions of 22cm × 22cm × 26cm (including the neck).
• Hardware: The system utilizes high-performance iDS U3-3881LE-AF global shutter cameras with liquid lenses, providing four optical DOF per camera (focus, exposure, white balance, gain). Motion is actuated by MG90S servos for the cameras and larger DC servos for the neck, all controlled by a Teensy 4.0 microcontroller. Two IMUs (BNO085) rigidly attached to the camera frames provide rotational state feedback at 60 Hz.
• Saccade Control Strategy: To address the difficulty of modeling complex kinematic offsets between motor axes and optical centers, the authors propose a data-driven saccade control method. Instead of explicit kinematic modeling, the system utilizes a calibration process where a ChArUco target is imaged at discrete pan and tilt motor steps.
  • Homography Mapping: The system computes homographies between calibration images to map the center of any image in the set to the pixel location of a target in the current view.
  • Interpolation: Since motor states are sampled discretely, a bilinear interpolation process is used to estimate motor values for unseen views based on the four closest calibration samples.
  • Execution: This approach allows the system to assign motor values directly to pixel locations without requiring precise axis alignment assumptions.

Key Contributions

1. DIJIT Platform: The introduction of an open-source robotic head that is, to the authors' knowledge, the first to combine a human-like baseline with full three rotational DOF per camera and three DOF for the neck, alongside four optical DOF per camera.
2. Novel Saccade Method: A fast calibration and control method that maps camera orientation to motor values using homography and interpolation. This method requires less than three minutes for calibration per camera and does not rely on the assumption that motor axes are perfectly aligned with optical axes.
3. Performance Benchmarking: A comprehensive evaluation of DIJIT's saccadic performance against both human data and other robotic systems in the literature.

Experimental Results
The authors evaluated DIJIT's saccadic performance through 175 trials involving binocular saccades to ArUco markers with varying eccentricities.

• Accuracy: The mean saccade error was 1.17° for the left camera and 1.14° for the right camera. After a corrective saccade (triggered when the primary error exceeded 1°), the error reduced to 0.93° and 0.95°, respectively. These values are comparable to reported human saccade errors (which range from 0.6° to 2.89° depending on conditions).
• Speed: DIJIT achieves saccade speeds of up to 600°/s, attaining approximately 85% of the peak human saccade speed (up to 700°/s).
• Repeatability: In a repeatability test of 100 trials, the standard deviation of landing point errors was low (e.g., 0.36° horizontally for the left camera), indicating high consistency.
• Comparison: DIJIT outperforms or matches other robotic heads in terms of speed and accuracy, despite using lower-resolution motor encoders than some competitors. The authors note that while some competing methods report errors in pixels, converting these to angular errors places DIJIT's performance favorably against systems like COG, OREO, and iCub.

Significance and Claims
The paper claims that DIJIT provides a unique platform for active vision research, enabling the study of human-like eye-head interactions and the comparison of biological versus computational visual strategies. The authors emphasize that the system is affordable, open-source, and replicable, making it accessible for cognitive science and neuroscience research. The significance of the work lies in its ability to bridge the gap between biological vision properties and robotic implementation, offering a system that matches human performance metrics in range, speed, and accuracy without the need for complex, error-prone kinematic modeling. The authors state that the device is currently being used to explore differences between human vision and current computer vision methods and is planned for deployment in autonomous systems to support wheelchair-bound individuals.