INRoL
Interactive & Networked Robotics Laboratory
Department of Mechanical & Aerospace Engineering
Seoul National University
MECHANICS/CONTROL
Online 3D Edge Reconstruction of Wiry Structures from Monocular Image Sequences
Three-dimensional (3D) reconstruction of wiry structures from vision suffers from thin geometry, lack of texture, and severe self-occlusions. To address these challenges, we propose an online 3D edge reconstruction framework that uses monocular image sequences to reconstruct the wiry structures by employing edges constructed from points as primitives of the representation. Specifically, we construct sparse 3D points via a robust maximum a posteriori (MAP) inference and use them to generate edges with beliefs, which is updated in a Bayesian fashion. We experimentally validate our framework using real-world wiry objects and demonstrate a manipulation task using the reconstruction, showing the potential to be easily utilized for subsequent robotic tasks. (RA-L 2023)
Uncertain Pose Estimation during Contact Tasks using Differentiable Contact Features
For many robotic tasks, accurate uncertain object pose estimation is key, usually requiring optimal fusion of geometry and sensor data. Previous solutions mostly rely on sampling-based or learning methods, facing efficiency and generalizability issues. We offer a novel differentiable framework for efficient and accurate uncertain pose estimation during contact. We introduce a new, adaptable geometric definition that allows differentiable contact features. Using a bi-level approach and gradient-based optimization, we efficiently solve for the uncertain pose. Tests show our framework outperforms existing methods. (RSS 2023)
Modular and Parallelizable Multibody Physics Simulation via Subsystem-Based ADMM
We introduce a new multibody physics simulation framework using subsystem-based structure and the Alternating Direction Method of Multipliers (ADMM). The main hurdle in simulating complex systems with many degrees of freedom is handling numerous coupled constraints and large matrices. We tackle this by dividing the multibody into subsystems and refactoring the dynamics equation accordingly. Using ADMM and a novel variable splitting scheme, we create a fast, scalable, and modular algorithm with good convergence and consistency. Performance evaluations demonstrate its advantages over existing algorithms. (ICRA2023)
Differentiable Dynamics Simulation Using Invariant Contact Mapping and Damped Contact Force
The gradient of typical differentiable simulation is uninformative for two reasons: 1) non-smoothness in contact dynamics not considered properly, and 2) excessive local minima generated from the smoothing procedure. To tackle this issue, we first propose differentiable contact dynamics with an invariant contact set and coordinate differentiation using a signed distance function (SDF). Also, to eliminate the undesirable jittering caused by the smoothing procedure, which induces extra local minima, and to achieve a smooth and informative gradient, we further endow our framework with a novel damped contact model. Various optimization problems are implemented to demonstrate the usefulness and efficacy of our differentiable framework. (ICRA 2023)
Large-Dimensional Multibody Dynamics Simulation Using Contact Nodalization and Diagonalization
We propose a novel numerical multibody simulation framework COND based on contact nodalization (with virtual nodes), and contact diagonalization (inter-contact and axes decoupling). This enables us to solve the surrogate dynamics for each iteration-loop in a one-shot/parallelized manner, and circumvent large-size/dense matrix factorization/multiplication, thereby, significantly speeding up the simulation time with improved convergence property. Performance and properties of our proposed simulation framework are also demonstrated and experimentally-validated for various large-dimensional/multi-contact scenarios including deformable objects. (T-RO)
Fast/Accurate Contact-Intensive Tight-Tolerance Simulation
We propose fast and accurate simulation framework for contact-intensive tight-tolerance robotic assembly tasks. The key components are: 1) data-driven contact clustering explicitly trained for physical accuracy and able to accommodate non-convex geometry; 2) contact solving precisely/robustly enforces physics of contact both for contact nodes and object configurations; 3) contact detection with neural network, parallelizable and fast computable even for complex shape objects with no exhaust pair-wise test; and 4) time integration with PMI (passive mid-point integration (Kim et al. (2017))) improves overall simulation accuracy, stability, and speed. We implement our proposed framework for two representative contact-intensive tight-tolerance tasks: peg-in-hole assembly and bolt-nut assembly, and validate its speed and accuracy against real experimental data. We also compare its performance with other physics engines and manifest its robustness via haptic rendering of virtual bolting task. (Arxiv)
Parallelized Iterative Simulation of Flexible Cable Manipulation
We propose a novel real-time physically-accurate simulator for flexible cable manipulation. We discretize the cable into multiple rigid links, each with NCP contact model and compliant coupling; and partition the cable into a number of subsystems. We then formulate the inter-subsystem consistency constraint as a certain analytical condition; and solve each subsystem in parallel with the contact model together with this consistency condition in an iterative manner, achieving both the speed and the accuracy of the simulation. Experimental validation/demonstration are also performed to show the theory. (ICRA21 Best Manipulation Paper Award Finalist)
Real-Time Simulation of Robotic Snap Connection Process
We propose a novel real-time physically-accurate simulation framework for snap connection process. We design our simulation to fully exploit peculiarities of process (small/smooth deformation, stiff connector, segmented contact) by adopting: 1) linear FEM modeling adequate to deal with small deformation while providing much faster speed as compared to nonlinear FEM; 2) reducing the dimension by balanced model reduction (BMR) and segmentation of the snap connector FEM model; 3) parallelized data-driven collision detection. Experimentally verified simulations are also performed to show the efficacy of our proposed simulation framework. (IROS2021)
Real-Time Accurate Simulation with Multi-Contact
Fast, accurate and stable simulation with multi-contact and tight-tolerance is crucial to data-driven approaches, e.g., deep reinforcement learning. Current robot simulators (e.g., ODE, Vortex, Bullet, etc.) do not provide this capability. For this, we propose a novel data-driven contact clustering based on the interaction network and trained with real experimental data. Combining this with our PMI (passive mid-point integrator, IJRR17), we could attain real-time, experimentally-validated simulation of peg-in-hole and bolting tasks with multi-contact and very tight tolerance, all impossible for other current simulators. (ICRA19)
Sim-to-Real Transfer of Tight-Tolerance Bolting Tasks
We propose a novel sim-to-real (S2R) framework for bolting tasks with tight tolerance and complex contact geometry. S2R is desirable for cost and safety,, yet, that of assembly task rare due to the lack of multi-contact simulator. We implement S2R transfer of nut tightening policy which is adaptive to uncertain bolt positions. For this, we develop a new multi-contact simulator, which adopt the configuration-space abstraction w/ fast/stable passive midpoint integrator (PMI). Sampling-based motion planning and LQT are used for nominal controller to be compliant and avoid local optima, whereas RL is used as a high-level controller to adapt to the uncertainty. (IROS2020)
Large-Scale Dual-Arm System on Flexible Support
We design a control law for the height-task teleoperation system, consisting of two KUKA LWRs, an actuated stage and a 10-meter telescoping mast (system developed by KAERI). Control objective is to allow end-effector to precisely track (unpredictable) human command, while suppressing mast vibration. To ensure robustness while handling with large system DOF (27), passive decomposition and passivity-based control techniques are adopted. Further, onboard sensing is attained attaching IMUs to each mast segment and utilizing data-driven POD/MAP estimation technique (ICRA19).
Under-Actuated Tendon-Driven Robot
We present a novel design framework for general under-actuated tendon-driven (UATD) robots to mimic desired free motion while maintaining posture during contact operations. The key enabler is stiffness decomposition, which allows us to decompose UATD robot into actuated and un-actuated space, thereby, allowing us to attain compliant free motion, while minimizing un-actuated space deformation. Design optimization is also proposed, which automatically provides active/passive tendon routing, joint spring, pre-tension, etc. (ICRA2018)
Arm-Stage System on Vertical Beam
Simultaneous trajectory tracking and vibration stabilization control of manipulator-stage system on vertical flexible beam, which will be used for operation in height. Euler-Bernoulli and Lagrange dynamics formulation are used with certain boundary conditions to model the system. Coordinate transformation and passive decomposition are used to decompose the total 7-DOF dynamics (4-DOF for arm+stage and 3-DOF for vibration) into vibration+stage and tracking objective. Passivity-based control with damping injection control is then design to these decomposed dynamics. EF-tracking experiment is performed to show the efficacy, for which the typical null-space based tracking control becomes unstable. (IROS2017)
PBC of Nonholonomic Systems
Passivity-based control (PBC) has been widely used for robotic manipulators with its superior robustness as compared to feedback linearization due to its exploitation of open-loop nonlinear dynamics rather than cancel them out. This passivity-based approach, yet, has been surprisingly missing for nonholonomic mechanical systems. In this work, we present passivity-based stabilization control for a certain class of nonholonomic systems. For this, we establish passive configuration decomposition (PCD) and propose passivity-based time-varying and switching controls. We also manifest when PCD is possible; and also establish equivalence of the proposed controls with kinematic controllability. (T-RO2017)
Passive Decomposition Theory
Passivity is a fundamental property of mechanical systems with close connection with Lyapunov control synthesis. At the same time, task can often be described by holonomic map h(q) (e.g., grasping shape). For this, we develop the theory of passive decomposition, i.e., we can decompose open-loop Lagrange robot dynamics into: (1) shape system, describing h(q)-dynamics; (2) locked system, describing system behavior with h(q) fixed; and (3) passive coupling between them. Locked and shape systems individually inherit Lagrange structure and passivity, greatly facilitating control design. Passive decomposition has been extended to Riemannian manifold and nonholonomic systems. (T-RO2010, T-AC2013)
Control of Under-Actuated UAVs
Quadrotor UAV is an under-actuated system, that is, its-DOF is 6 evolving in SE(3), yet, its actuation is only 4-DOF (i.e., 4 rotors). We investigated the issue of control of this under-actuated quadrotor UAVs. In particular, we utilize backstepping control technique to overcome its under-actuation and also combine it with adaptation to address inertial uncertainty. We also extend this backstepping framework to distributed control of multiple quadrotors with their information flow constrained by a balanced graph. (Automatica2012, RAS2014, Best Paper Award IAS2012)