Research

Research Topics

The two main research directions of the Laboratory of Automation and Robotics, deformable objects and human–robot collaboration, and the other topics we work on.

Main research directions

What we are working on

Diagram of a safety filter for dual-arm manipulation of an elastic rod: conditions for shape control (equilibrium regularity, shape authority), physical safety limits (bending, twist, collision), and the filter correcting the robot commands with a rod model
01 · Main direction

Robotic Perception and Manipulation of Deformable Linear Objects

Deformable Linear Objects (DLOs) are a specific subset of Deformable Objects (DOs), i.e. objects that can change their shape when subjected to external forces. DLOs are characterized by a linear geometry and include examples such as wires, cables, strings, and ropes. Despite being common in both domestic and industrial environments, DLOs pose a significant challenge for robotic applications. Indeed, the manufacturing and assembly industries that work with wires and wiring harnesses still rely heavily on human labor. The main challenges that robots face when dealing with these objects are: (1) the lack of any specific shape or feature that distinguishes them from other objects; (2) the vast number of possible configurations they can assume; and (3) their highly nonlinear and complex dynamics, which are generally difficult to model and predict.

At LAR, we are actively researching approaches to effectively perceive DLOs using deep learning methods, while also developing smart strategies to enable the automatic generation of training datasets. For the robotic manipulation, our focus is on combining both model-based and learning-based approaches, with the aim of addressing some of their individual limitations.

Recent work extends this line to language-grounded perception, multi-view tracking, topology learning for the dual-arm manipulation of multiple cables, and complete robotic cells for connector assembly and wiring-harness manufacturing. We also reviewed the field in a survey and in a two-part tutorial on shape servoing of deformable objects.

Key publications

All results and papers Related publications
02 · Main direction

Shared Autonomy, Human–Robot Collaboration and Robotic Hands

Robots that work side by side with people must understand their intentions and share control with them in a natural way. At LAR, we develop interfaces based on surface electromyography (sEMG) that decode muscle activity to control robotic hands and assistive devices, with regression approaches that need little or no labelled data.

We study shared-autonomy schemes in which the robot assists the user, for instance by regulating the grip strength of a myocontrolled robotic hand or by supporting the manipulation of deformable objects, and we use sEMG-based estimates of muscle co-contraction and stiffness to make kinesthetic teaching and teleoperated programming by demonstration more intuitive. This line builds on the lab’s long experience in mapping human hand motions to robotic hands and in the design and control of anthropomorphic robot hands, such as the UB Hand.

Key publications

All results and papers Related publications

Other research topics

Also in the lab

03

Robotic Manipulation of Liquids in Open Containers

Sloshing is the oscillatory motion of a liquid inside a moving, partially filled container. When liquid-filled containers are moved between stations in processing or packaging lines, or carried on a robotic tray without being grasped, sloshing has to be limited to avoid spillage, product loss and slow, overly cautious motions.

We develop computationally cheap models that estimate the sloshing height during general motions, such as 3D translations combined with a rotation about the vertical axis (SCARA motions) or about a horizontal axis (tilting), based on equivalent mass-spring-damper and pendulum models that need no experimental tuning and no external sensors. Building on these models, we plan time-optimal anti-sloshing trajectories for several containers transported at once, compute closed-form container tilting that compensates sloshing for motions with a prescribed yaw, and use model predictive control for non-prehensile transport that guarantees both contact stability and a bound on the wave height. To validate the models, we measure the maximum sloshing height with RGB action cameras; the measurement tools and the experimental data sets are publicly released.

  • Sloshing-height estimation
  • Anti-sloshing trajectory planning
  • Tilting compensation
  • Non-prehensile transport (MPC)
  • Vision-based sloshing measurement

Key publications

Related publications
A Franka arm moves away from an approaching hand; next to it, the robot's distance-field model, the point cloud of the hand and the repulsive directions
04

Real-Time Collision Avoidance with Distance Fields

Robots that share their workspace with people and moving objects must react to obstacles within milliseconds, without abandoning the task they are executing. We embed distance fields directly into task-priority control: avoidance of obstacles and self-collisions sits in the same hierarchy as end-effector tracking, joint limits and singularity avoidance, and is activated smoothly only when it is needed.

The robot links are modelled as signed distance fields built from Bernstein polynomials, which give a continuous geometry and closed-form gradients for the repulsive actions; low-rank tensor decompositions make their batched evaluation on the GPU fast and memory-efficient. The environment is sensed with depth cameras and laser scanners, as point clouds or as a GPU voxel map with its Euclidean distance transform. The framework runs on fixed-base and mobile manipulators; on mobile manipulators, the redundancy of the holonomic base is also used to steer repetitive trajectories away from singular configurations.

  • Signed distance fields
  • Task-priority control
  • GPU acceleration
  • Self-collision avoidance
  • Mobile manipulators
  • Singularity avoidance

Publications

The 3D-printed monolithic compliant wrist in four poses of flexion and deviation
05

Mechatronic Design: Compliant Wrists, Grippers, Actuators and Tactile Sensors

Many of the devices we use are designed and built in the lab. A recent line concerns cable-driven compliant wrists for upper-limb prostheses and robots: two degrees of freedom are obtained with cross-axis flexural pivots, printed as a single monolithic part that is lightweight, frictionless and free from backlash, and that matches the range of motion of the human wrist (±80° of flexion/extension, ±40° of ulnar/radial deviation). Fast virtual prototypes based on finite element analysis let us tune the controller before the physical device is built.

Other designs include a tendon-driven three-fingered gripper for autonomous underwater vehicles, an additively manufactured flexible endoscope driven by antagonistic twisted string actuators, and tactile sensors: a low-cost soft skin with an embedded IMU that localizes contacts on large surfaces, and a physics-based calibration that turns an optoelectronic tactile array into a force sensor for grasping cables.

  • Compliant mechanisms
  • Prosthetic wrists
  • Virtual prototyping
  • Tendon-driven grippers
  • Twisted string actuation
  • Tactile sensing

Publications