💻 computer science

Admittance-Based Motion Planning with Vision-Guided Initialization for Robotic Manipulators in Self-Driving Laboratories

This paper proposes a motion-planning framework for robotic manipulators in self-driving laboratories that combines vision-guided initialization via structured planar pose estimation with an embedded admittance controller to enable adaptive, compliant, and human-interactive manipulation.

Original authors: Shifa Sulaiman, Tobias Jensen, Francesco Schetter, Simon Bøgh

Published 2026-02-10

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Shifa Sulaiman, Tobias Jensen, Francesco Schetter, Simon Bøgh

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are working in a high-tech laboratory where robots are doing all the heavy lifting, mixing chemicals, and moving delicate glassware. Now, imagine one of those robots is moving a tray of test tubes, and you suddenly need to nudge it out of the way to reach a tool.

In a traditional setup, the robot is like a stiff, unyielding statue. If you push it, it will either resist you with all its might (which is dangerous) or it might "break" its programming and error out because you interfered with its path.

This paper introduces a smarter way for robots to work in these "Self-Driving Laboratories." They’ve given the robot two superpowers: "Smart Eyes" and "Soft Touch."

1. The "Smart Eyes" (Vision-Guided Initialization)

Before the robot can move, it needs to know what it’s looking at. Instead of a human typing in exact coordinates, the researchers gave the robot a vision system.

The Analogy: Think of this like a person walking into a dark room with a flashlight. The robot uses its cameras to scan the room, finds a specific object (like a book with a pattern on it), and "draws" a mental map of exactly where that object is and how it’s tilted. It’s like the robot is saying, "Aha! I see the target, and I know exactly how to grab it."

2. The "Soft Touch" (Admittance-Based Motion Planning)

This is the core of the paper. The researchers implemented something called Admittance Control.

The Analogy: Imagine the difference between a wooden broomstick and a pool noodle.

A wooden broomstick is rigid. If you try to bend it, it fights back or snaps. This is a traditional robot.
A pool noodle is compliant. If you push it, it moves with your hand. This is the "Admittance" robot.

The researchers programmed the robot to act like that pool noodle. They created a mathematical "buffer" that tells the robot: "Follow your planned path, but if you feel a physical force (like a human hand), don't fight it. Yield to it, move out of the way, and once the pressure is gone, gently go back to what you were doing."

Why does this matter?

In a busy lab, things are unpredictable. A human might need to step in, or a piece of equipment might be slightly out of place.

By combining these two things, the researchers created a robot that can:

See a target on its own (The Vision).
Plan a path to grab it (The Intelligence).
Dance out of the way if a human touches it (The Compliance).

The Results

The researchers tested this in a computer simulation and then with a real, heavy-duty mobile robot.

In the simulation: The robot was nearly perfect, moving smoothly like a well-choreographed dancer.
In the real world: Even when they pushed the robot with significant force, it didn't panic or glitch. It "yielded" to the push, moved aside, and then—just like a polite assistant—returned to its task once the path was clear.

In short: They are building robots that aren't just "automated machines," but "collaborative partners" that can see what they are doing and feel when they are being touched.

Technical Summary: Admittance-Based Motion Planning with Vision-Guided Initialization

1. Problem Statement

Self-driving laboratories (SDLs) aim to automate scientific discovery by integrating robotics with AI. However, most current robotic frameworks in these environments lack compliant interaction capabilities. Standard rigid motion planning makes robots unable to adapt to unexpected environmental changes, such as contact with delicate laboratory equipment or physical intervention by a human operator. There is a critical need for a system that can both accurately localize objects using vision and respond safely to external forces through compliant control to ensure safety and adaptability in human-robot collaborative settings.

2. Methodology

The authors propose a dual-layered framework that combines vision-based initialization with an admittance-based control scheme.

Vision-Guided Initialization:
To establish the initial target configuration, the system uses a structured planar pose estimation algorithm. It employs Scale-Invariant Feature Transform (SIFT) for feature extraction and FLANN for descriptor matching. A Homography matrix ( $H$ ) is computed via RANSAC to model the planar transformation of textured objects. The object's 3D pose is then refined using depth data from an RGB-D camera, constructing an orthonormal basis $(\vec{i}, \vec{j}, \vec{k})$ to define the object's orientation in space.
Admittance-Based Motion Planning:
Instead of traditional rigid trajectory execution, the framework embeds an admittance controller directly into the execution loop. The controller treats the manipulator as a virtual mass-spring-damper system, governed by the dynamic equation:
$M_d\ddot{x}(t) + B_d\dot{x}(t) + K_d(x(t) - x_d(t)) = F_{ext}(t)$
This allows the robot to map external forces ( $F_{ext}$ ) to a Cartesian reference velocity ( $\dot{x}_{ref}$ ), which is then converted to joint space via the manipulator Jacobian. The final commanded trajectory is a blend of the nominal planned path and the admittance-induced displacement, allowing the robot to "yield" to external forces and subsequently resume its path.
Experimental Platform:
The framework was tested on a mobile manipulator consisting of a Clearpath Ridgeback omni-directional base, a UR5e collaborative arm, and a Robotiq Hand-E gripper, equipped with Intel RealSense cameras and LiDAR.

3. Key Contributions

Compliance-Integrated Motion Planning: A novel scheme that embeds admittance control into trajectory execution, enabling the robot to be physically redirected by humans and safely resume tasks.
Vision-Guided Pose Estimation: A robust pipeline for detecting and localizing textured planar objects to provide real-time target configurations for motion planning.
Integrated Framework for SDLs: A proof-of-concept that bridges the gap between high-level perception (vision) and low-level safe interaction (admittance control) specifically tailored for laboratory environments.

4. Results

The framework was validated through both simulation (ROS/Gazebo) and physical experiments.

Vision Performance: The system achieved a grasping success rate of 93.7% in experiments. Pose estimation error was kept within $\pm$ 0.61 cm, with a detection latency of 71 ms (~13 FPS), ensuring real-time responsiveness.
Admittance Control Performance:
- Robustness: The robot successfully responded to external forces up to 18 N without instability.
- Trajectory Fidelity: In experiments, the maximum trajectory deviation was 1.02 cm, with a recovery time of 1.21 s.
- Human-Robot Collaboration: The controller demonstrated a 34% reduction in human effort during physical interaction, proving its ergonomic benefits.
Comparison: While simulation yielded slightly higher precision (e.g., 0.43 cm pose error vs. 0.61 cm in experiments), the experimental results closely mirrored the theoretical mass-spring-damper model.

5. Significance

This research provides a foundation for safer and more autonomous robotic integration in scientific research. By enabling robots to be "force-aware," the framework allows for seamless human-robot collaboration, where operators can intervene without triggering emergency stops or causing damage. This is a vital step toward fully autonomous self-driving laboratories where robots must navigate cluttered, unpredictable, and delicate environments. Future work aims to extend this to transparent and deformable objects, further increasing the autonomy of SDLs.