Technical Summary: CoLI – A Reproducible Platform for Continuum Robot Learning

Problem Statement

Continuum robots possess high degrees of freedom, compliant structures, and inherent safety, making them ideal for unstructured and confined environments such as minimally invasive surgery and human-robot interaction. However, their adoption in research and practical learning-based applications is hindered by three primary barriers:

1. Fabrication Complexity: Traditional designs rely on complex assembly of custom-fabricated rigid and flexible components, leading to high costs, low reproducibility, and significant dimensional errors.
2. Control Challenges: The continuously deformable morphology and actuation redundancy make accurate kinematic modeling difficult. Conventional task-space control often suffers from workspace mismatches and fails to effectively exploit the robot's redundant degrees of freedom.
3. Lack of Learning-Ready Infrastructure: There is a scarcity of accessible hardware platforms that support scalable data collection and imitation learning, as existing open-source solutions often lack standardized pipelines for continuum robots.

Methodology

The authors present CoLI, a novel open-source continuum robot system designed to lower barriers for hardware fabrication and control. The methodology encompasses four integrated components:

1. Monolithic Multi-Material 3D Printing

The robot is fabricated as a single, continuous compliant structure using multi-material Fused Deposition Modeling (FDM).

• Structure: The arm consists of a flexible backbone printed from TPU 95A and 16 rigid spacer disks printed from PLA. These form 15 serial joints enabling continuous bending.
• Design: The backbone features planar elliptical arcs to provide compliance and elastic restoring forces, while tangent segments ensure robust connections to the disks.
• Process: Using a Bambu Lab H2D printer, the entire body is printed in a single 29-hour run (0.2 mm layer height). Post-processing requires only 15 minutes to remove support material, eliminating manual assembly of the backbone and disks.
• Customizability: The design allows for rapid iteration of joint counts, dimensions, and morphologies (e.g., slender or tower-like variants) without retooling.

2. Actuation Mechanism

The robot employs a tendon-driven architecture with independent actuation.

• Configuration: 15 joints are grouped into three sets, each controlled by three tendons routed through guide holes in the spacer disks.
• Actuators: Tendons are driven by Dynamixel XC430-T240BB-T motors arranged at the base. This setup allows for direct tension regulation via motor rotation, eliminating the need for manual tensioning and enabling load sharing.

3. Isomorphic Teleoperation

To facilitate intuitive control and data collection without relying on complex kinematic models, the system uses a leader–follower teleoperation interface.

• Isomorphism: The "leader" manipulator is mechanically identical to the "follower" robot but utilizes lower-reduction-ratio motors (Dynamixel XL330-M288-T) for better backdrivability.
• Mapping: Control is achieved through a direct, joint-wise mapping where the leader's motor encoder positions are mapped to the follower's target positions. This eliminates the need for inverse kinematics, providing a singularity-free, actuator-aligned control scheme.
• Calibration: Both robots are calibrated to a zero-reference state using cylindrical sleeves and PWM-driven tendon pretension.

4. Imitation Learning Pipeline

The system is integrated into the LeRobot framework to support autonomous control.

• Data Collection: The system records motor positions (relative to zero offsets) and multi-view camera data (end-effector and external) at 30 Hz.
• Policy Training: An Action Chunking with Transformers (ACT) model is trained on the collected demonstrations. The policy uses a ResNet-18 backbone and a 4-layer transformer encoder to predict robot actions from proprioceptive and visual inputs.

Key Contributions

The paper outlines four main contributions:

1. Fabrication Paradigm: A monolithic multi-material 3D printing approach enabling single-run, reproducible manufacturing of continuum robots, significantly reducing assembly complexity.
2. Isomorphic Teleoperation: A control framework that establishes a direct correspondence between operator inputs and robot actuators, facilitating intuitive full-body control and high-quality demonstration collection for imitation learning.
3. Integrated Platform: A learning-ready continuum robot system bridging hardware and embodied AI pipelines via the LeRobot framework.
4. System Validation: Comprehensive characterization of the robot's structure, teleoperation accuracy, and autonomous manipulation capabilities, establishing a reproducible baseline for the community.

Experimental Results

The platform was evaluated through hardware characterization and manipulation tasks:

• Mechanical Properties:

  • Force-Deformation: The backbone exhibits an approximately linear force-deformation relationship with a slope of ~3.2 N/mm.
  • Workspace: The reachable workspace is approximately a 430 mm diameter sphere.
  • Payload: The robot can reliably support a dynamic payload of up to 1 kg while maintaining controllability. Performance degrades noticeably at 1.2 kg.
  • Durability: The system operated for over 15 hours without structural damage or performance degradation.

• Teleoperation Performance:

  • Accuracy: The follower robot achieved a mean positional tracking error of 8.2 mm (3.3% of arm length) across 127 configurations.
  • Latency: The system demonstrated a settling-time latency of 158 ms at a 30 Hz control frequency, confirming suitability for high-fidelity demonstration collection.
  • Dexterity: Operators successfully completed complex tasks, including picking and placing objects and retrieving items from a confined 90 mm diameter tube.

• Imitation Learning:

  • Three autonomous tasks were evaluated: Object Placement, Switch Toggling, and Planar Pushing.
  • Success Rates: The robot achieved success rates of 100% (12/12) for object placement, 83% (10/12) for switch toggling, and 83% (10/12) for planar pushing.
  • The policies demonstrated robust error recovery and retry behaviors in challenging scenarios.

Significance and Claims

The paper claims that CoLI significantly lowers the barriers to entry for continuum robot research by providing a reproducible, learning-ready platform. By unifying monolithic fabrication, isomorphic teleoperation, and standardized learning pipelines, the system addresses the critical lack of accessible hardware and intuitive control interfaces that have previously limited data-driven manipulation research in this domain.

The authors position this work not as a final solution but as a foundational step to accelerate algorithmic development and systematic benchmarking. They acknowledge limitations, including the build volume constraints of 3D printing on arm length and the high cost of servo motors, and propose future work focused on modular designs, lower-cost actuation, and richer sensing modalities. The ultimate goal is to foster community-driven progress in continuum robot learning through standardized, open-source resources.