Autumn 2020 Colloquium

Organizers: Karthik Desingh, Dieter Fox, Maya Cakmak, Siddhartha S. Srinivasa

Explainable Robotics
Nima Keivan (Amazon) 01/10/2020
Reinforcement Learning for Real-World Robotics
Mrinal Kalakrishnan (X) 01/17/2020
Lazy Belief-Space Task and Motion Planning for Robots Acting in Partially-Observable Environments
Caelan Garrett (MIT) 01/24/2020
Observing, Learning, and Planning Fine-Grained Manipulation Activities: A Systems Perspective
Greg Hager (Johns Hopkins University) 02/14/2020
Acceptance over Ignorance - How to embrace uncertainty in robotic manipulation
Jeannette Bohg (Stanford University) 02/21/2020
Allen School Robotics Overview
Dieter, Sidd, Maya, Joshua, Byron 10/09/2020

Abstract: Faculty will introduce their lab members, give a quick overview of their research, and talk about working remotely.

[Cancelled event] Marita Cheng 10/16/2020
Robotics-inspired biomechanics: modeling and predicting complex responses to exoskeletons
Michael Rosenberg (University of Washington) 10/23/2020

Abstract: Assistive devices, such as orthoses, prostheses, and exoskeletons are commonly used to help individuals with motor impairments – such as children with cerebral palsy or stroke survivors – improve gait and walk more efficiently. However, predicting how specific individuals will adapt their gait pattern to novel device designs remains challenging. Time-intensive experimental device optimization is the most effective approach to tuning assistive devices for an individual. Modeling approaches ubiquitous in biomechanics often rely on assumptions about an individual’s physiology, which are often invalid for individuals with motor impairments, limiting the accuracy and utility of model predictions. Drawing inspiration from methods proposed to study dynamical systems and control bipedal robots expands our ability to quantitatively customize assistive devices in silico, without requiring prior knowledge of an individual’s physiology or motor control. In this talk, I discuss the data-driven and data-plus-physics-driven approaches that our lab has used to model and predict gait with ankle exoskeletons. First, I will show how we used phase-varying models to predict responses to ankle exoskeleton torque without knowledge of an individual’s physiology. I will then discuss our use of template models of locomotion and sparse regression to identify statistically-supported and interpretable reduced-order dynamics describing how an individual regulates locomotion. Data-driven modeling for gait analysis gait may generalize to other types of interventions, enabling individualized quantitative rehabilitation protocols to improve treatment efficacy and the individual’s mobility.

Biography: Michael is a PhD candidate in the University of Washington Mechanical Engineering Department, working under Dr. Kat Steele. Michael’s research uses data-driven methods to inform ankle exoskeleton design by modeling and predicting subject-specific changes in gait in response to varying ankle exoskeleton mechanical properties. In addition to his primary dissertation research, Michael has mentored undergraduate students on exoskeleton-focused research projects ranging from muscle synergy analysis to predictive musculoskeletal modeling. Michael completed his MS in mechanical engineering at the University of Washington and his BS in mechanical engineering at North Carolina State University. He is a recipient of an NSF Graduate Research Fellowship and a Predoctoral Training Fellowship from the Institute for Translational Health Sciences.

Exploring Novel Input Techniques for Human-Robot Interaction: Is Manual Input Always the Best Choice?
Momona Yamagami (University of Washington) 10/23/2020

Abstract: During robot teleoperation where continuous guidance of robot movement is needed, such as during robot-assisted surgery or robot navigation of complex terrain, manual control of such robots using joysticks, mice, and handles is the norm. Understanding if and when emerging technologies like muscle or brain interfaces provide advantages over conventional manual interfaces requires modeling the human as part of the human-robot system. My research demonstrates the potential of muscle (electromyography, EMG) interfaces as an alternative input technique for humans to continuously control robots. I demonstrate how human input can be quantified as a response to error (feedback control) and future position (feedforward control) and how this method can be used to quantify human performance during a one-dimensional trajectory-tracking task. The results suggest that people controlling mechanical systems and people with limited movement from stroke may benefit from using muscle interfaces as an alternative input method to manual interfaces, and highlight the need for exploration of novel input techniques outside of traditional manual interfaces.

Biography: Momona Yamagami is a PhD candidate in the ECE department working with Profs. Sam Burden and Kat Steele on modeling and enhancing human-robot interactions using novel input techniques. Her research focuses on how continuous interactions can be improved for people with and without limited movement.

ME themed Robotics Overview
Ashis Banerjee, Xu Chen, Sawyer Fuller, Jeffrey Lipton 10/30/2020
Blending Model-Predictive Control and Value Function Approximation for Efficient Reinforcement Learning
Mohak Bhardwaj (University of Washington) 11/06/2020

Abstract: Model-Predictive Control (MPC) is a powerful tool for controlling complex, real-world systems that uses a model to make predictions about future behavior. For each state encountered, MPC solves an online optimization problem to choose a control action that will minimize future cost. This is a surprisingly effective strategy, but real-time performance requirements warrant the use of simple models. If the model is not sufficiently accurate, then the resulting controller can be biased, limiting performance. In this work, we present a framework for improving on MPC with model-free reinforcement learning (RL). Our key insight is to view MPC as tracing out a series of local Q-function approximations. We can then blend each of these Q-functions with value estimates from RL to systematically trade-off errors between the two sources. We present a theoretical analysis that shows how error from inaccurate models in MPC and value function estimation in RL can be balanced, and test the efficacy of our approach on challenging high-dimensional manipulation tasks with biased models in simulation. We demonstrate that our approach can obtain performance comparable with MPC with access to true dynamics even under severe model bias and is more sample efficient as compared to model-free RL.

Biography: Mohak Bhardwaj is a Ph.D. CSE student at the University of Washington advised by Byron Boots. His research primarily focuses on developing theory and algorithms for scalable and efficient robot learning with a specific focus on the intersection of reinforcement learning, model-predictive control and motion planning.

Learning Reactive Robot Motion Policies with Control-theoretic Guarantees
Anqi Li (University of Washington) 11/06/2020

Abstract: Traditionally, modeling and control techniques have been regarded as the fundamental tools for studying robotic systems. Although they can provide theoretical guarantees, these tools make limiting modeling assumptions. Recently, learning-based methods have shown success in tackling problems that are challenging for traditional techniques. Despite its advantages, it is unrealistic to directly apply most learning algorithms to robotic systems due to issues such as sample complexity and safety concerns. In this line of work, we aim at making robot learning explainable, sample efficient, and safe by construction through encoding structure into policy classes. In particular, we focus on a class of structured policies for robotic problems with multiple objectives. The complex motions are generated by combining simple behaviors given by Riemannian Motion Policies (RMPs). It can be shown that the combined policy is stable if the individual policies satisfy a class of control Lyapunov conditions, which can imply safety. Given such a policy representation, we learn policies with such structure so that formal guarantees are provided. To do so, we keep the safety-critical policies, e.g. collision avoidance and joint limits policies, as fixed during learning. We can also make use of the known robot kinematics. We show learning with such structure effective on a number of learning from human demonstration tasks and a simple reinforcement learning task.

Biography: Anqi Li is a PhD student in the Paul G. Allen School of Computer Science and Engineering at the University of Washington working with Byron Boots. Her research focuses on providing theoretical guarantees for robot learning using tools from control theory.

Contact-less manipulation of mm-scale objects via ultrasonic levitation
Jared Nakahara (University of Washington) 11/13/2020

Abstract: Although general purpose robotic manipulators are becoming more capable at manipulating various objects, their ability to manipulate millimeter-scale objects are usually limited. On the other hand, ultrasonic levitation devices have been shown to levitate a large range of small objects, from polystyrene balls to living organisms. By controlling the acoustic force fields, ultrasonic levitation devices can compensate for robotic manipulator positioning uncertainty and control the grasping force exerted on the target object. The material agnostic nature of acoustic levitation devices and their ability to dexterously manipulate millimeter-scale objects make them appealing as a grasping mode for general purpose robots. In this work, we present an ultrasonic, contact-less manipulation device that can be attached to or picked up by any general purpose robotic arm, enabling millimeter-scale manipulation with little to no modification to the robot itself. This device is capable of performing the very first phase-controlled picking action on acoustically reflective surfaces. With the manipulator placed around the target object, the manipulator can grasp objects smaller in size than the robot's positioning uncertainty, trap the object to resist air currents during robot movement, and dexterously hold a small and fragile object, like a flower bud. Due to the contact-less nature of the ultrasound-based gripper, a camera positioned to look into the cylinder can inspect the object without occlusion, facilitating accurate visual feature extraction.

Biography: Jared is an ECE graduate student in Joshua Smith’s Sensor System’s lab working on developing acoustic levitation devices that can control and manipulate objects within the context of robotics, medical devices, and scientific tools. He received his B.S. in Electrical Engineering from the University of Washington in 2018. Jared is interested in acoustic levitation devices that take advantage of high frequency sound to impart force on objects allowing these devices to move objects and overcome the force of gravity. The benefits of this manipulation method include contactless grasping, mechanical manipulator error compensation, grasping without object occlusion and safe manipulation of fragile and living organisms. These attributes lend itself to applications such as robotics, medical devices, scientific tools and advanced manufacturing systems.

Competitive Physical Human-Robot Game Play
Boling Yang (University of Washington) 11/13/2020

Abstract: Competition is one of the most common forms of human interaction, yet competitive interaction has rarely been discussed in the context of Human Robot Interaction (HRI). There has indeed been a large focus in HRI on cooperative interaction, such as human-aware motion planning, object handover actions, and collaborative manipulation. Conversely, the absence of studies in competitive robot interaction may be due to anxieties concerning the actions of robots whose interest do not necessarily align with our own. However, these fears should not prohibit us from considering positive impacts that competitive-HRI can yield, such as providing the participants with motivation, inspiring their potential, and more. In this work, we aim to motivate research in competitive human-robot interaction (competitive-HRI) by discussing how human users can benefit from robot competitors from a psychological perspective. We then examine the concepts from game AI that we can adopt for competitive-HRI. Based on these discussions, we propose a robotic system that is designed to support future competitive-HRI research. A human-robot fencing game is also proposed to evaluate a robot's capability in competitive-HRI scenarios. Finally, we will present the initial experimental results and discuss possible future research directions.

Biography: Boling Yang is a Ph.D. student in The Paul G. Allen School of Computer Science and Engineering at the University of Washington advised by Joshua Smith. His current research interests include robotic manipulation and perception, human robot interaction, and reinforcement learning.

Learning to Navigate in the Real World without a Metric Map
Xiangyun Meng (University of Washington) 11/20/2020

Abstract: Traditional approaches for robot navigation requires meticulous reconstruction of environments via Simultaneous Localization and Mapping (SLAM). While SLAM is highly effective, it requires expensive or bulky hardware such as laser scanners and high-resolution cameras for accurate reconstruction. This is in contrast to our everyday experience, where we can effortlessly navigate in large-scale environments without building a metric map. This powerful visual navigation capability of humans and animals has drawn significant interests recently thanks to the advancements in deep learning. While a wide variety of spatial representations and learning methods have been proposed to tackle this problem, few works have been shown to work on real robots due to their unrealistic assumptions and poor scalability. In this talk, I will present a minimalistic system that enables a real robot to visually navigate in the real world without a metric map. In particular, we do not assume the availability of a global coordinate system or robot poses, nor do we assume noise-free actuation or obstacle-free environment . The system comprises two components: a robust low-level controller that models robot dynamics and avoids obstacles, and a high-level visual topological memory for mapping and planning. We show that with our system, a robot can i) navigate robustly with the presence of motion constraints, actuation noise and obstacles; ii) build a compact spatial memory through adaptive experience sparsification and iii) generalize to real robots while trained only in simulation. Finally, I will share our recent work on leveraging behaviors to improve robustness and memory efficiency of our system.

Biography: Xiangyun Meng is a fourth year Ph.D. student supervised by Prof. Dieter Fox. His research lies in the intersection between perception, planning and control in the setting of visual navigation.

Multimodal Trajectory Prediction via Topological Invariance for Navigation at Uncontrolled Intersections
Junha Roh (University of Washington) 11/20/2020

Abstract: We focus on decentralized navigation among multiple non-communicating rational agents at uncontrolled intersections, i.e., street intersections without traffic signs or signals. Avoiding collisions in such domains relies on the ability of agents to predict each others’ intentions reliably, and react quickly. Multiagent trajectory prediction is NP-hard whereas the sample complexity of existing data-driven approaches limits their applicability. Our key insight is that the geometric structure of the intersection and the incentive of agents to move efficiently and avoid collisions (rationality) reduces the space of likely behaviors, effectively relaxing the problem of trajectory prediction. In this paper, we collapse the space of multiagent trajectories at an intersection into a set of modes representing different classes of multiagent behavior, formalized using a notion of topological invariance. Based on this formalism, we design Multiple Topologies Prediction (MTP), a data-driven trajectory-prediction mechanism that reconstructs trajectory representations of high-likelihood modes in multiagent intersection scenes. We show that MTP outperforms a state-of-the-art multimodal trajectory prediction baseline (MFP) in terms of prediction accuracy by 78.24% on a challenging simulated dataset. Finally, we show that MTP enables our optimization-based planner, MTPnav, to achieve collision-free and time-efficient navigation across a variety of challenging intersection scenarios on the CARLA simulator.

Biography: Junha Roh is a Ph.D. student in Paul G. Allen School of Computer Science and Engineering at the University of Washington, advised by Dieter Fox and Ali Farhadi. His research focuses on vision language navigation and autonomous driving.

Differentiating Shape with Respect to Motion
William Agnew (University of Washington) 12/04/2020

Abstract: Object shape provides a strong prior over dynamics and behavior, enabling simulation of object dynamics when coupled with physics simulators. However, accurately inferring object shape from video in realistic, cluttered scenes remains an open problem. Existing 3D reconstruction algorithms often produce shapes that, while visually similar, yield dramatically different dynamics. Further, there is no general way to make inferences about object shape based on observed dyanmics. In this paper we propose a method for making object shape differentiable with respect to dynamics. Our method allows for training reconstruction neural networks to optimize how closely the simulated dynamics of reconstructions match observed dynamics, making them produce reconstructions more useful for simulation and planning. In addition, our method allows for optimizing estimated object shape to match observed dynamics, allowing agents to take advantage of highly informative motion cues.

Biography: William is a Ph.D. student in Computer Science at University of Washington. He is advised by Pedro Domingos and Sidd Srinivasa and supported by an NDSEG Fellowship. His research focuses on developing human priors for reinforcement learning, with projects in object oriented reinforcement learning.

Leveraging Post-Hoc Context for Faster Learning in Bandit Settings with Applications in Robot-Assisted Feeding
Ethan Gordon (University of Washington) 12/04/2020

Abstract: A successful robot-assisted feeding system requires bite acquisition of a wide variety of food items. It needs to adapt to changing user food preferences under uncertain visual and physical environments. Different food items in different environmental conditions may require different manipulation strategies for successful bite acquisition. Therefore, a key challenge is to handle previously-unseen food items with very different success rate distributions over strategy. Previous work showed that the problem can be represented as a linear contextual bandit on visual information. However, this setting discards additional haptic context that can be acquired during or after the execution of the action.

Biography: Ethan is a third-year PhD student in the Personal Robotics Lab at the University of Washington. His current research interests are in robotics and AI with a focus on online learning for physics-based deformable object manipulation and the application of assistive feeding. He has previously done work on varifocal systems at Oculus and integrated photonics at Princeton. In his spare time, he enjoys skiing, a Cappella, Dungeons and Dragons, and ham radio.

Acting with Language
Mohit Shridhar (University of Washington) 12/11/2020

Abstract: Language is a powerful tool for humans and robots. A tool that helps communicate intentions, and also to think about the world in terms of abstract concepts. In this talk, I'll present two frameworks for using language with embodied agents: (1) communicating goals and instructing agents on how to achieve these goals, and (2) using language as a medium for abstract reasoning to capture priors like locations of objects, object affordances, and pre-conditions. Representing goals and plans with language allows agents to generalize better to new tasks and environments where lower-level modalities (like vision and proprioception) often struggle. With these frameworks, we provide a playground to study the interaction between language grounding, state-estimation, and long-horizon planning.

Biography: Mohit is a Ph.D student in the Robotics State Estimation (RSE) Lab at the University of Washington. He is advised by Dieter Fox. His research interests are in grounded language learning for robotics and vision applications.

Discovering and Segmenting Unseen Objects for Robot Perception
Chris Xie (University of Washington) 12/11/2020

Abstract: In order to function in unstructured environments, robots need the ability to recognize unseen objects. This is a challenging perception task since the robot needs to learn the concept of "objects" and generalize it to unseen objects. In this report, we investigate different methods for learning such perception systems by exploiting different visual cues and learning from synthetic data. We first develop a novel neural network architecture, PT-RNN, that leverages motion cues by casting the problem as object discovery via foreground motion clustering from videos. This network learns to produce pixel-trajectory embeddings such that clustering them results in segmenting the unseen objects into different instance masks. Next, we introduce UOIS-Net, which separately leverages RGB and depth for unseen object instance segmentation. UOIS-Net is able to learn from synthetic RGB-D data where the RGB is non-photorealistic, and provides state-of-the-art unseen object instance segmentation results in tabletop environments, which are common to robot manipulation.

Biography: Chris Xie is currently a PhD student in the Paul G. Allen School of Computer Science and Engineering at the University of Washington, advised by Dieter Fox. From 2016 to 2019, he was a National Defense Science and Engineering Graduate (NDSEG) Fellow. His research interests include applying machine learning to solve robot perception problems.