IRIS: Learning-Driven Task-Specific Cinema Robot Arm for Visuomotor Motion Control

Qilong Cheng, Matthew Mackay, Ali Bereyhi

Abstract

Robotic camera systems enable dynamic, repeatable motion beyond human capabilities, yet their adoption remains limited by the high cost and operational complexity of industrial-grade platforms. We present the Intelligent Robotic Imaging System (IRIS), a task-specific 6-DOF manipulator designed for autonomous, learning-driven cinematic motion control. IRIS integrates a lightweight, fully 3D-printed hardware design with a goal-conditioned visuomotor imitation learning framework based on Action Chunking with Transformers (ACT). The system learns object-aware and perceptually smooth camera trajectories directly from human demonstrations, eliminating the need for explicit geometric programming. The complete platform costs under $1,000 USD, supports a 1.5 kg payload, and achieves approximately 1 mm repeatability. Real-world experiments demonstrate accurate trajectory tracking, reliable autonomous execution, and generalization across diverse cinematic motions.

Introduction

Professional cinematic robotic systems enable dynamic and repeatable camera motions that surpass human capabilities. However, their widespread adoption is hindered by the high cost (often exceeding $50,000) and the technical complexity of programming industrial-grade platforms. Existing low-cost 3D-printed alternatives typically fail to meet the reach, payload, and structural requirements for professional cinematography.

We introduce the Intelligent Robotic Imaging System (IRIS), a task-specific 6-DOF manipulator designed to bridge this gap. IRIS is a sub-$1,000 platform that leverages a goal-conditioned visuomotor imitation learning framework to perform professional-grade camera moves. By co-designing lightweight hardware with a transformer-based controller, IRIS achieves sub-millimeter repeatability and produces motion smoother than the human experts who trained it.

System Architecture and Mechatronics

The IRIS project adopts a "full-stack" co-design approach, integrating hardware development with advanced policy learning.

The IRIS system pipeline, spanning mechatronics design, data collection, low-level control, and high-level policy learning.

Hardware Design

To maintain a low cost while ensuring high performance, IRIS utilizes Quasi-Direct Drive (QDD) actuators (Unitree GO-M8010-6 BLDC motors). These motors provide high torque density and backdrivability, which is essential for kinesthetic teaching. A key design innovation is the reduction of distal inertia; by employing a differential wrist and a belt-driven elbow, heavy motors are positioned closer to the base. The structural frame consists of 3D-printed components reinforced with 25mm x 25mm carbon fiber tubes for increased rigidity.

Exploded view of the IRIS 6-DOF manipulator, highlighting the differential joint, carbon fiber structural links, and actuator placement.

Control Framework

The system operates on a dual-rate control loop. A low-level 200Hz ROS-based impedance control loop manages the physical dynamics and joint states. Above this, a high-level policy runs at 10Hz, predicting 15-step "chunks" of future joint positions. This "Action Chunking" approach ensures smooth transitions and temporal consistency in the camera's path.

Learning-Driven Motion Control

The "brain" of IRIS is based on Action Chunking with Transformers (ACT), adapted to be goal-conditioned. Instead of following a pre-programmed path, the robot is provided with a "target image" representing the desired final framing of the shot.

Handling Multimodality

Human cinematic motion is inherently stylistic and multimodal—there are many ways to execute a "push-in" shot. To capture this, the researchers integrated a Conditional Variational Autoencoder (CVAE). The CVAE allows the model to learn various "styles" of motion from human demonstrations, ensuring the robot can generalize across different creative intents.

Data Collection

Training data was gathered through kinesthetic teaching, where a human expert manually guides the robot in a zero-torque mode. The resulting dataset comprises 132 episodes (approximately 14,000 training clips) focusing on "push-in" (dolly-in) maneuvers toward objects like a coffee cup, including scenarios with and without obstacles.

Experimental Results and Insights

IRIS was evaluated on its ability to reach target framing with high precision and smoothness.

  • Precision and Payload: Despite its $1,000 price point, the system achieved approximately 1mm repeatability and successfully stabilized a 1.5 kg payload, suitable for high-end mirrorless cameras.
  • The "Perceptual Filter" Effect: A significant finding was that the trained policy produced trajectories with a jerk of 0.61 $m/s^3$, compared to the 3.64 $m/s^3$ recorded from human experts. The neural network effectively filtered out high-frequency human jitters, resulting in superior cinematic quality.
  • Closed-Loop Robustness: The ACT-CVAE policy achieved a 90% success rate in reaching target framing, significantly outperforming classical geometric planners like RRT* (10% success). The learning-based approach successfully compensated for the physical flex and calibration nuances of the 3D-printed hardware that classical models failed to account for.
  • Absolute Control: Ablation studies confirmed that predicting absolute joint positions was vital for success. Incremental control strategies failed (0% success) due to error accumulation in the low-cost hardware.

Conclusion and Future Work

IRIS demonstrates that a task-specific, learning-driven approach can deliver professional-grade cinematic performance on consumer-grade hardware. While current limitations include mechanical flexion during high-torque moves and a limited repertoire of shots, the system provides a foundation for accessible robotic cinematography. Future work will focus on expanding the motion library to include pans, tilts, and complex multi-goal sequences.

Summary updated

Abstract

Robotic camera systems enable dynamic, repeatable motion beyond human capabilities, yet their adoption remains limited by the high cost and operational complexity of industrial-grade platforms. We present the Intelligent Robotic Imaging System (IRIS), a task-specific 6-DOF manipulator designed for autonomous, learning-driven cinematic motion control. IRIS integrates a lightweight, fully 3D-printed hardware design with a goal-conditioned visuomotor imitation learning framework based on Action Chunking with Transformers (ACT). The system learns object-aware and perceptually smooth camera trajectories directly from human demonstrations, eliminating the need for explicit geometric programming. The complete platform costs under $1,000 USD, supports a 1.5 kg payload, and achieves approximately 1 mm repeatability. Real-world experiments demonstrate accurate trajectory tracking, reliable autonomous execution, and generalization across diverse cinematic motions.
BibTeX
@misc{cheng2026iris,
  author = {Qilong Cheng and Matthew Mackay and Ali Bereyhi},
  title = {{IRIS: Learning-Driven Task-Specific Cinema Robot Arm for Visuomotor Motion Control}},
  year = {2026},
  eprint = {2602.17537},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.17537},
}
Figure