Here you can watch the recorded talks of the IROS 2021 workshop

Sam Burden

Sam Burden (University of Washington, USA)

Title: On infinitesimal contraction analysis for mechanical systems subject to unilateral constraints

Abstract: A system is contractive if all trajectories converge toward one another –- it is infinitesimally contractive if this global dynamical property can be inferred from local dynamics. Contractive systems enjoy strong properties, e.g. any equilibrium or periodic orbit is globally asymptotically stable, and they are easy to control optimally and compose robustly. This talk considers infinitesimal contractivity in the class of mechanical systems subject to unilateral constraints that model contact-rich dynamics in robot locomotion and manipulation. We will start by describing our recent results on contraction analysis for hybrid systems, which builds on our prior work metrizing hybrid state spaces and approximating local dynamical properties in nonsmooth and hybrid systems. Then we will focus on mechanical systems by considering representative examples with linear unilateral constraints. Finally, we will discuss how robots should be designed and controlled to produce contractive dynamics, and conclude by detailing the benefits of contractivity for robust composition and optimization- or learning-based control

Watch Sam Burden’s talk

Bernard Brogliato

Prof. Bernard Brogliato (INRIA, France)

Title: On the control of juggling nonsmooth Lagrangian systems

Abstract: Juggling systems encompass “true” juggling systems, but also hand-manipulation of objects with persistent contact and with or without sliding at fingers' tips, manipulation by pushing, tapping, walking and running biped robots dynamics, jumping robots, crawling robots (where the center of gravity dynamics plays the role of the object in both cases), kinematic chains with joint clearance, etc. Thus they make a large class of underactuated nonsmooth Lagrangian systems. As the above simple example shows there may be two main control approaches: through impacts (true jugglers, running robots, manipulation by tapping), or in persistent contact (in which case friction may be the prominent effect, like in crawling and walking robots). Many different works have been dedicated to the control of this class of systems, but it seems that a systematic analysis is still missing. In this talk some results will be reviewed with some open questions.

Watch Bernard Brogliato’s talk

Kevin Green

Kevin Green (Oregon State University, USA)

Title: Embracing Ground Uncertainty in Control of Agile Bipedal Robots

Abstract: Legged robots will encounter a range of unanticipated ground terrain, and cannot rely entirely on perception, which is often noisy, late, and wrong. The foundation of robust legged locomotion is dynamic behavior, motions and reflexes that are specifically designed to take unanticipated ground variations in stride. Observations from biology and robot optimization studies emphasize that particular leg swing behavior and a significant foot velocity at contact with the ground can provide the desired robustness. We have shown highly successful learned control policies on Cassie using first-principle goals from biomechanics and reduced-order robot models. These control policies are extremely robust and are even able to reliably scale flights of stairs with no perception information.

Watch Kevin Green’s talk

Kaveh Akbari Hamed

Prof. Kaveh Akbari Hamed (Virginia Tech, USA)

Title: Hierarchical and Nonlinear Feedback Control of Legged Robots: From Hybrid Systems to Planning and Robust Control

Abstract: The past few years have seen an accelerated effort to design rehabilitation and emergency response robots and to develop robots with human and animal traits. Legged locomotion is extremely important in this advancement. The study of legged locomotion has been motivated by the desire to allow people with disabilities to walk and to assist humans in hazardous environments. Legged robots that can perform at this level do not yet exist, and part of what is holding back their development and deployment is adequate feedback control theory for nonlinear and hybrid dynamical models of locomotion subject to unilateral constraints and underactuation. While rapidly advancing technologies enable the design of increasingly sophisticated legged robots, the pace of hardware innovations is exceeding the ability of computational control algorithms. In particular, there is a fundamental gap in knowledge of scalable algorithms required for motion planning as well as resilient feedback control synthesis of high-dimensional hybrid models arising from agile and dexterous legged locomotion in complex environments. In this talk, we present a systematic approach to design hierarchical and nonlinear feedback controllers, based on hybrid systems theory and optimization, to enable real-time planning and robust stabilization of a multitude of locomotion patterns for legged machines. Our approach provides a systematic and computationally attractive solution to design centralized as well as decentralized/distributed controllers that can robustly stabilize different locomotion patterns for single- and multiagent legged machines. In this talk, we will show how to design hierarchical and nonlinear controllers for sophisticated models of legged locomotion based on geometric control techniques, model predictive control (MPC), control Lyapunov functions (CLFs), control barrier functions (CBFs), and distributed control. The higher level of the proposed control schemes is developed based on a real-time optimal control problem of reduced-order models subject to feasibility conditions. We will then investigate the stability properties of these reduced-order models for the steering problem. To bridge the gap between reduced- and fullorder models of locomotion, quadratic programming (QP)-based nonlinear controllers subject to stability, safety, and feasibility (i.e., CLFs and CBFs) are developed at the lower level of the proposed control schemes to impose the full-order dynamics to asymptotically track the optimal reduced-order trajectories. The power of these algorithms will be finally demonstrated in designing robust stabilizing nonlinear controllers for 1) locomotion of advanced quadrupedal robots, 2) locomotion of quadrupedal robots with robotic tails, and 3) cooperative locomotion of multiagent legged robots that collaboratively transport objects.

Watch Kaveh Hamed’s talk

Yildirim Hurmuzlu

Prof. Yildirim Hurmuzlu (Southern Methodist University, USA)

Title: A new method in solving multi-contact impact problems of kinematic chains

Abstract: In this presentation I will focus on the multi-impact problems of planar kinematic chains (particle and rigid body) with external objects. These problems are frequently arise in the fields of Robotics and Biomechanics. Such problems can be described as a kinematic chain with one of its ends striking an external surface while the remaining ends resting on other surfaces. This type of problem involves complementarity relationships between the normal velocities and impulses at the contacting ends. I will go over a method of solving such problems that does not require a search algorithm and it is efficient and produces a unique solution. In addition, I will cover the topic of critical configurations of chains during the impact process. Critical configurations of particle and rigid body chains are important in various applications where the impulse wave generated by impact gets blocked before it reaches a contacting end.

Watch Yildirim Hurmuzlu’s talk

Aaron Johnson

Prof. Aaron Johnson (Carnegie Mellon University, USA)

Title: Transparency and Proprioceptive Contact Localization

Abstract: In order for robots to interact more naturally with unstructured environments, they need to be able to bump into things and not fall over or damage their surroundings. But many robots have high inertia and slow contact sensing, making it hard for them to have a light touch. To quantify how small a total impulse a robot can have with the environment, we present a new transparency metric that captures the closed-loop response to an impact. This metric combines the initial impulse, a compression phase before contact is detected, and then a reaction phase as the robot works to remove the contact. Analyzing this metric we see that many designs that have recently gained renewed interest, in particular low gear ratio or direct drive actuation, score more favorably than traditional robot designs. One key to driving down the total impulse is fast and accurate contact detection. In this talk we also present a new proprioceptive contact localization approach using velocity constraints that is well suited to transparent robot designs. The method is less sensitive to model errors than torque-based approaches and provides an instantaneous estimate of contact location along a link. We show how these results apply to both legged locomotion (with a Ghost Robotics quadruped) and manipulation (with a CMU DDHand).

Watch Aaron Johnson’s talk

Karen Liu

Prof. Karen Liu (Stanford University, USA)

Title: Optimizing Physical Contact for Motor Control: Turning a Challenge into a Solution

Abstract: Leveraging physical contacts to interact with our surroundings is an essential skill to achieve any physical task, but contact-rich, dynamically changing environments often create significant challenges to autonomous robotic locomotion and manipulation. While there exists computationally tractable contact models to aid the development of robust control policies, the discontinuities inherent in the contact phenomenon introduce non-differentiability in the equations of motion, rendering traditional approaches to optimal control ineffective. In this talk, I will show that, with intelligent contact control and planning algorithms, the challenge of handling contact can become a solution. The first part of the talk focuses on a learnable physics simulator that models contact behaviors based on a small amount of data via adversarial learning. The second part of the talk focuses on a differentiable physics engine that efficiently computes accurate gradients through contact constraints formulated as a Linear Complementarity Program. Lastly, I will mention some work we have done in the area of data-driven haptic perception for robot-assisted dressing tasks.

Watch Karen Liu’s talk

Hae-Won Park

Prof. Hae-Won Park (Korea Advanced Institute of Science and Technology, South Korea)

Title: Optimization-based Design, Control, and Estimation for Dynamic Legged Locomotion

Abstract: Dynamic legged locomotion is a challenging robotic task that naturally incorporates dynamic impact events. Recently, various locomotion capabilities of legged robots that embrace dynamic impact events have been achieved such as dynamic walking, high-speed running, and jumping over obstacles. Among many technological advances in the field of legged robots, a novel actuator design, optimization-based control and estimation strategies have played a central role in the attempts toward the use of legged robots in real-world scenarios. In this talk, I will talk about diverse aspects of dynamic legged locomotion robots including actuator design, model predictive control, and state estimation, all based on optimization, and how these technologies have been integrated and deployed to hardware platforms to achieve more robust and dynamic performances in legged robots.

Watch Hae-Won Park’s talk

Michael Posa

Prof. Michael Posa (University of Pennsylvania, USA)

Title: Perspectives on multi-contact robotics: deep learning, impact-invariant control, and modeling non-uniqueness

Abstract: Impacts between robot and world, integral to dynamic walking, running, and high-speed manipulation, are notoriously difficult to predict or control. Simultaneous contact is particularly challenging to manage. For example, multiple contacts can produce scenarios where non-unique outcomes can arise from rigid-body assumptions. However, due to the inability to predict these outcomes, this possibility has largely been ignored in planning and control methods. I will present a single differential inclusion, and accompanying sampling-based algorithm, to capture the entire set of possible outcomes. Next, I will discuss an approach to control during these impact events. To deal with the uncertainty prevalent when striking the ground, common approaches use heuristics to blend control strategies or reduce gains. Here, based on a representation of uncertainty, we develop an impact-invariant strategy that reduces sensitivity to impacts while maintaining some control authority and demonstrate this strategy for jumping with the Cassie robot. I’ll conclude the talk with highlights of our work on learning models for multi-contact motion. Standard deep learning approaches introduce biases that clash with the physical properties of friction and impacts. Our approach, ContactNets, is explicitly designed to represent and learn through discontinuities and is dramatically more accurate and data efficient.

Watch Michael Posa’s talk