Aug. 27 |
Jeff Mahler |
University of California, Berkeley |
4-5pm |
250, Sutardja Dai Hall, UC Berkeley |
|
Click to read abstract Rapid and reliable robot grasping of a wide variety of objects remains a Grand Challenge for robotics due to sensor noise, imprecise control, and partial observability. Deep neural networks trained on datasets of human-labeled or self-supervised grasps can be used to rapidly plan grasps across a diverse set of objects, but data collection is tedious and performance may asymptote with training dataset size. To reduce data collection time, I propose to generate synthetic training datasets of millions of 3D point clouds and robot grasps using geometric models of grasp success and image formation. In this talk I present the Dexterity-Network (Dex-Net), a framework for generating datasets by analyzing mechanical models of contact forces and torques under stochastic perturbations across thousands of 3D object CAD models. I describe generative models for training policies to lift and transport objects from a tabletop or cluttered bin using a parallel-jaw (two-finger) or suction cup gripper. I explore methods for learning robust policies that transfer from simulation to reality and that can decide which gripper to use. To substantiate the method, I describe thousands of experimental trials on a physical robot which suggest that deep learning on synthetic Dex-Net datasets can be used to rapidly plan successful grasps across a diverse set of novel objects with a 95% success rates on heaps of 25 objects. |
Sept. 10 |
Katie Campbell |
University of Illinois at Urbana-Champaign |
4-5pm |
250, Sutardja Dai Hall, UC Berkeley |
|
Click to read abstract Autonomous systems, such as self-driving cars, are becoming tangible technologies that will soon impact the human experience. However, the desirable impacts of autonomy are only achievable if the underlying algorithms can handle the unique challenges humans present: People tend to defy expected behaviors and do not conform to many of the standard assumptions made in robotics. To design safe, trustworthy autonomy, we must transform how intelligent systems interact, influence, and predict human agents. In this work, we’ll use tools from robotics, artificial intelligence, and control to explore and uncover structure in complex human-robot systems to create more intelligent, interactive autonomy. In this talk, I'll present on robust prediction methods that allow us to predict driving behavior over long time horizons with very high accuracy. These methods have been applied to intervention schemes for semi-autonomous vehicles and to autonomous planning that considers nuanced interactions during cooperative maneuvers. I’ll also present a new framework for multi-agent perception that uses people as sensors to improve mapping. By observing the actions of human agents, we demonstrate how we can make inferences about occluded regions and, in turn, improve control. Finally, I’ll present on recent efforts on validating stochastic systems, merging learning and control, and implementing these algorithms on a fully equipped test vehicle that can operate safely on the road. |
Sept. 17 |
Mark Palatucci |
ANKI |
4-5pm |
250, Sutardja Dai Hall, UC Berkeley |
|
Click to read abstract For the past several decades, consumer applications of robotics have been more science fiction than reality. However, recent developments in deep-learning, cloud AI, and plummeting prices of both computation and sensing have created the necessary components for a rapidly growing consumer robotics industry to finally emerge. In this talk, I’ll discuss the evolution of Anki from 3 Ph.Ds and a kitchen table prototype, to a global company that has quickly become the 2nd largest producer of consumer robots in the world. I’ll share many of the successes and challenges of producing robots at million+ unit scale, and the important trends that will impact both academia and industry. I’ll talk about the importance of emotion and character for building a great user experience, and some surprising findings about human-robot interaction. I’ll also discuss Anki’s unique “bottom’s up approach" to robotics, and show how with an increasingly complicated series of low-cost mass-market robots, we’ve created a virtuous cycle that’s driving growth in the industry. Lastly, I'll discuss Vector, a new robot Anki just announced that is an important step towards a future with useful, emotive, robot companions for every home. |
Sept. 24 |
Jur van den Berg |
University of Utah |
4-5pm |
250, Sutardja Dai Hall, UC Berkeley |
N/A |
Click to read abstract In this talk, Jur will share his passion for the specific application of autonomous vehicle technology to moving freight on highways, as it meaningfully downscopes the technical challenges of autonomous mobility in multiple ways, while at the same time allowing for a compelling business case. He will explain in detail what the specific technical challenges are to letting trucks drive themselves safely on highways, and what it takes to get a self-driving product on the road, for real. |
Oct. 1 |
Zico Kolter |
Carnegie Mellon University |
4-5pm |
250, Sutardja Dai Hall, UC Berkeley |
|
Click to read abstract Although wildly successful, deep learning systems are also extremely brittle; this is evidenced for example by the widespread possibility of adversarial attacks, specially crafted inputs meant to fool deep classifiers. This talk will discuss our recent work in develop deep classifiers that are provably robust to (certain classes of) perturbation attacks. Our methods work by considering a convex relaxation of the "adversarial polytope", the set of last-layer activations achievable under some norm-bounded perturbation of the input, and using these to derive very efficient methods for computing (and then minimizing) upper bounds the adversarial loss that can be suffered under such attacks. The method leads to some of the largest verified networks of which we are currently aware, including a convolutional MNIST classifier with a provable bound of 3.7% error under L_infinity perturbations of size epsilon=0.1. I'll relate our work to similar ongoing directions, and also discuss the main challenges that it faces: the task of scaling to significantly larger networks, e.g. at ImageNet scale, and the task of better characterizing the "correct" set of perturbations that we would like to be robust to. I'll also connect the work to efforts in proving properties about deep networks in other settings, such as control and general verification. |
Oct. 8 |
Sertac Karaman |
Massachusetts Institute of Technology |
4-5pm |
250, Sutardja Dai Hall, UC Berkeley |
|
Click to read abstract The robotics research community has long relied on the developments in computing technology for consumer electronics in order to integrate complex, performant algorithms into robotics applications. However, the slowdown in the improvements in general computing systems have led the consumer electronics industry to develop application-specific chips from the ground up for specific applications, ranging from face detection to augmented reality. This shift in computing technology presents substantial opportunities and significant technical challenges for next-generation embedded systems for robotics. In the first part of this talk, we discuss ultra-high-throughput embedded systems for robotics applications. We envision fast, agile vehicles equipped with high-rate, high-resolution sensing and powerful embedded computing systems. We discuss high-dimensional control design using compressed continuous computation with tensor decompositions, high-throughput visual-inertial navigation using feature selection with sub-modular optimization, and other problems at the intersection of motion planning and perception. We propose new algorithms with provable guarantees on performance. We demonstrate some of the algorithms in an autonomous drone racing scenario, reaching more than 6.7m/s (roughly 15mph) in a room that is 7m (roughly 20 feet) in size, in a virtual reality environment.In the second part of the talk, we discuss ultra-low-energy embedded systems for robotics. We envision insect-size vehicles with complex perception and decision-making capabilities. We argue that, in order to enable vehicles at such scale, their computers must be designed from the ground up together with the algorithms. We discuss our latest application-specific integrated circuit (ASIC) for visual-inertial odometry with bundle adjustment. The new chip requires a mere 2 milliWatts on average, and can process camera images up to 170 frames per second. |
Oct. 15 |
Jerry Ding |
United Technologies Research Center |
4-5pm |
250, Sutardja Dai Hall, UC Berkeley |
|
Click to read abstract This talk provides an overview of autonomy and robotics research at the United Technologies Research Center (UTRC), from its past heritage in autonomous vehicle technology development for Sikorsky Aircraft, to its present day focus areas on human-robot collaboration for a broad range of commercial and aerospace applications within United Technologies Corporation (UTC). This includes a review of the role of UTRC in maturing and transitioning key technologies in the areas of autonomy architecture, planning, perception, and human-machine systems which led to successful flight demonstrations on full-sized rotorcraft platforms. The talk will also describe how similar capabilities continue to be carried forward within the context of the current robotics initiative at UTRC, with applications in the areas of advanced manufacturing, inspections, service, and maintenance. A selection of active research topics will also be presented, including hierarchical planning, collaborative autonomy, and human robot interaction. This is accompanied by concept demonstration results on a range of UAV, UGV, and robotic manipulation platforms. |
Oct. 22 |
David Gealy |
University of California, Berkeley |
4-5pm |
250, Sutardja Dai Hall, UC Berkeley |
|
Click to read abstract Recent advances in robot learning for control have shown increasing success for tasks in real-world human applications. However, we believe capable robots are still too expensive to allow for widespread consumer adoption for applications within elder care, and homes. We discuss considerations for a new design paradigm for low cost capable robotic manipulators, and follow through with a fully functional realization of this new paradigm. |
Oct. 29 |
Ian Mitchell |
University of British Columbia |
4-5pm |
250, Sutardja Dai Hall, UC Berkeley |
|
Click to read abstract Recent demonstrations of continuous state reachability in thousands of dimensions are impressive, but to maximize control authority while ensuring safety for human-in-the-loop systems we need not just to identify the existence of a safe trajectory, but to characterize the set of safe controls. In the first part of the talk I will describe a new algorithm for constructing under-approximations of robust controlled invariant or viable sets of uncertain linear systems using an efficient convex optimization. In the second part I will tackle a challenge for robust analyses: The a priori model uncertainty and hence safe control authority may be overly conservative. I will describe an algorithm which addresses this challenge by detecting and efficiently removing the effect of erroneous uncertainty at runtime, thereby improving control authority when it is safe to do so. The algorithms will be demonstrated on a nonlinear model of a quadrotor and automated delivery of anesthesia respectively. |
Nov. 5 |
Liam Pedersen |
Nissan |
4-5pm |
250, Sutardja Dai Hall, UC Berkeley |
|
Click to read abstract The promise of driverless vehicles to transform the urban mobility experience is contingent on them behaving well within the human driving and pedestrian ecosystem. Driverless vehicles must behave in a socially acceptable manner. This talk will unpack what this means and explore how the technologies for achieving this. |
Nov. 26 |
Kris Pister |
University of California, Berkeley |
4-5pm |
250, Sutardja Dai Hall, UC Berkeley |
N/A |
Click to read abstract Several decades ago in the early days of MEMS, micro robots seemed like they were just around the corner. The first IEEE MEMS conference was the ambitiously-named "Micro Robots and Teleoperators Workshop". But every piece of building micro robots turned out to be hard. Motors, mechanisms, sensing, computation, communication, and power all required significant improvement to support the creation of autonomous robots on a sub-centimeter size scale. After several decades of progress, it appears that we are finally at the point where we will be able to build such systems, making our own silicon insects and other robots with no natural analog. This talk will cover the status of our mesh-networked silicon ant, flea, and ionocraft efforts, with some hints at future needs for control, path planning, navigation, swarm control, etc. |
Dec. 3 |
Christopher D. Gill |
Washington University |
11am-12pm |
400 Cory Hall, UC Berkeley |
N/A |
Click to read abstract New advances in parallel real-time scheduling theory and concurrency platforms are enabling a new generation of cyber-physical systems, which can support a challenging combination of (1) significant computational demands, (2) stringent timing constraints, and (3) dynamic and substantial changes in how resources are allocated, at run-time. This talk will describe a series of recent advances in parallel real-time systems research, including both theoretical and practical results, and how each of them impacts the specific domain of real-time hybrid simulation, a cyber-physical approach to high-fidelity testing of structures at scale that is increasingly relevant for earthquake engineering and other related fields. |
Dec. 3 |
Alessandro Astolfi |
Imperial College London |
4-5pm |
250, Sutardja Dai Hall, UC Berkeley |
N/A |
Click to read abstract Several nonlinear control analysis and design problems, such as stability and stabilization problems, adaptive control problems, model reduction and observer design, optimal/robust control and game theory problems, rely upon the computation of solutions of partial differential equations (PDEs). These could be linear PDEs, such as those encountered in stability, stabilization, and adaptive control and observer design; nonlinear PDES, such as the Hamilton Jacobi PDEs; or coupled PDES, such as those arising in game theory and in mean field games. In this talk we present a method to replace PDEs with algebraic equations at the expense of sacrificing performance (in a measurable way). The method is shown to be sufficiently flexible to deal with a wide range of problems and to provide solutions which are more accurate than those obtained using series expansion methods or other numerical schemes for the solutions of PDEs. Examples in the areas of engine control, multi agent systems and state estimations illustrate the theory. |
Dec. 4 |
Martin Törngren |
Kungliga Tekniska Högskolan Royal Institute of Technology |
4-5pm |
540 A/B, UC Berkeley |
N/A |
Click to read abstract Automated driving is breaking new ground where "smart", connected and eventually collaborating machines are deployed for increasingly complex tasks in unstructured environments. This provides a wonderful stage with numerous opportunities and challenges, especially considering the societal impact and broad variety of applications, from automated machines in mines to SAE level 4 automated cars. This stage is representative of an ongoing technological shift, with similar trends in many other domains. However, current methodologies are not well prepared for such future Cyber-Physical Systems (CPS), requiring new systems and safety engineering approaches to be established. In this talk I will address limitations of existing engineering methodologies, and in particular those of safety engineering. Traditional safety engineering is focused on structured environments and risk reduction, which based on risk assessment and the definition of corresponding risk reducing measures. For CPS and so called functional safety, the approaches are to a large extent process-based and treat software as deterministic. Common safety engineering patterns include the use of "safety functions" that are separate from the nominal system and that are simple enough to be "certified" w.r.t. the identified risk level. Using this context, I will discuss the design and assurance of highly automated vehicles, elaborating a number of perspectives to the challenges and directions, emphasizing the role of controlled experimental environments, providing minimal performance requirements, and accelerating the development of safety engineering guidelines. I will describe our ongoing work on so called safety supervisor architectures and their design. A conceptual architectural design of a fault-tolerant autonomous driving intelligence will be presented, encompassing a nominal and a safety supervisor channel. I will discuss hazardous events, their sources, the design space in terms of redundancy and diversity, the division of responsibilities among the channels, when the supervisor should take over, and remaining open challenges. In architecting future automated vehicles it is essential that viewpoints and solution strategies from controls, computer science, computer engineering, AI/ML, safety and classical vehicle engineering are brought together. |
Dec. 10 |
Sonia Martinez |
University of California, San Diego |
4-5pm |
250, Sutardja Dai Hall, UC Berkeley |
N/A |
Click to read abstract Self-organization is a pervasive phenomenon in nature, which has inspired the development of multi-robot systems that can mimick their biological counterparts. As we consider larger groups of autonomous agents or swarms, new theoretical challenges appear that are associated with both their size and robotic-system limitations. In this talk, we outline recent work on two complementary problems related to the control of large swarms. First, we consider a deployment objective by which robots are to be shaped into certain density profile. Under the assumption that agents can obtain measurements of the local density, but do not have access to absolute position information, we propose a PDE-based feedback control strategy that includes the distributed computation of diffeomorphisms. Then we discuss how to handle complementary set of limitations such as lack of access to position information, noisy actuation, or optimal transport under distributed gradient flows. To finish, we present preliminary results on the identification of subsets of nodes that are critical for the performance of spatial consensus algorithms in large groups. |