Join us for an engaging and insightful seminar with Steve Cousins, the Executive Director of the Stanford Robotics Center, as he introduces the groundbreaking work being done at one of the world’s leading hubs for robotics innovation. In this talk, Steve will provide an overview of the Center’s mission and provide insight into the future of robotics.

Compliance is a physical property of motion that describes the elastic relationship brings force and motion variations. A suitable compliance profile brings robustness to robotic manipulation by handling uncertainties gracefully. In this talk, I will introduce two sets of methods for designing compliance control in manipulation tasks. I will first walk through manipulation robustness analytically, and show the role compliance control can play to improve it. With basic modeling information, the optimal control/motion plan can be computed efficiently. Then I will talk about how to learn a compliant manipulation policy directly from human demonstrations. We propose Adaptive Compliance Policy (ACP), a framework that learns to dynamically adjust system compliance both spatially and temporally for given manipulation tasks from human demonstrations.

Recent advances across multiple research fields are rapidly changing the way in which we develop autonomous systems. In this talk, I will discuss how space autonomy can benefit from the rise of foundation models. The discussion will focus on two perspectives. First, I will discuss how techniques that are traditional to the foundation model literature can be adapted for the purpose of reliable decision-making in space, with a focus on the application of Transformers for spacecraft trajectory optimization. Next, I will discuss the opportunities presented by pre-trained foundation models within future machine learning-based autonomy stacks for space applications, ranging from data curation to serving as reconfigurable automated reasoning modules within modular autonomy stacks, towards the goal of developing a broadly capable Space Foundation Model.

Living robots represent a new frontier in engineering materials for robotic systems, incorporating biological living cells and synthetic materials into their design. These bio-hybrid robots are dynamic and intelligent, potentially harnessing living matter’s capabilities, such as growth, regeneration, morphing, biodegradation, and environmental adaptation. Such attributes position bio-hybrid devices as a transformative force in robotics development, promising enhanced dexterity, adaptive behaviors, sustainable production, robust performance, and environmental stewardship. Nature’s musculoskeletal design can act as an inspiration for both artificial and living robots. We will explore recent advances in artificial electrohydraulic musculoskeletal robots, which employ electrohydraulic actuators to produce lifelike muscle contractions and adaptive motions, as demonstrated in our recent work published in Nature Communications. We will also discuss our breakthroughs in vision-controlled inkjet printing for robotics from our Nature paper, as well as xolographic biofabrication techniques for biohybrid swimmers presented at RoboSoft. Additionally, I’ll share insights from our computational optimization of musculoskeletal systems featured at Humanoids. Together, these projects showcase how musculoskeletal, bio-hybrid, and computational techniques are opening new frontiers in robotics interaction and manipulation.

ReachBot is a joint project between Stanford and NASA to explore a new approach to mobility in challenging environments such as martian caves. It consists of a compact robot body with very long extending arms, based on booms used for extendable antennas. The booms unroll from a coil and can extend many meters in low gravity. In rocky environments the booms are equipped with low-mass grippers that use spines for a secure grasp. The booms are strong in tension but vulnerable to buckling in compression or bending. Motion planning with ReachBot therefore has similarities to multifingered grasp planing -- instead of fingers that push, we have booms that pull. Given its very long reach, ReachBot has a large dexterous workspace that simplifies motion planning. However, the sequence of poses must also consider what happens if any grasp fails. In this talk I will introduce the ReachBot design and motion planning considerations, report on a field test with a single ReachBot arm in a lava tube in the Mojave Desert, and discuss future plans, which include the possibility of mounting one or more ReachBot arms equipped with wrists and grippers on a mobile platform – such as ANYMal. To learn more: http://bdml.stanford.edu/ReachBot

When will robots be able to clean my house, dishes, and take care of laundry? While we source labor primarily from automated machines in factories, the penetration of physical robots in our daily lives has been slow. What are the challenges in realizing these intelligent machines capable of human level skill? Isn’t AI advanced enough to replace many skills of humans? Unlike conventional robots, which are optimized mainly for position control with almost no adaptability, household tasks require a kind of 'physical intelligence' that involves complex dynamic interactions with overwhelming uncertainties. While advanced language models exemplify AI's prowess in data organization and text generation, a significant divide exists between AI for virtual and physical applications. In this conversation, we'll delve into the cognitive biases that often lead us to underestimate this technological gap.

Fast dynamics, nonlinearities, and tedious tuning are just a few of the many reasons why we are interested in leveraging learning for control—challenges that are ubiquitous in robotics and other physical machines. In this talk, we will explore the problem of learning controllers through three paradigms, organized from general to structured learning problems: deep reinforcement learning, automatic imitation learning from optimal control, and auto-tuning via Bayesian optimization. I will highlight some of our recent results addressing key challenges faced in practice, such as enhanced uncertainty quantification for improved data efficiency and reliability in model-based reinforcement learning, as well as parameter-adaptive approximate model predictive control for imitation learning without retraining. By discussing these advancements alongside applications—demonstrated through hardware experiments on unicycle robots, quadcopters, and cars—I aim to develop an understanding of the potential of these paradigms in both research and current practice.

As we deploy robots in increasingly dynamic and unstructured environments with data-driven policies, the need to be able to make guarantees on the reliability and safety of these systems keeps growing. In this talk, I will present two perspectives on uncertainty quantification. First, I will present a conformal prediction-based framework for making in-distribution guarantees on the safety of a learned perception and planning system. Next, I will present a planning framework for out-of-distribution guarantees using coherent risk measures. I will provide the experimental validation of these methods on ground robots for navigation and showcase applications for subterranean search and rescue. Finally, I will present future directions and challenges in attaining reliable autonomy under distribution shifts.

While there has been major investment in developing large-scale robot learning algorithms, achieving true autonomy remains a wide-open research question. A key ingredient towards this goal is a robots ability to generalize to unseen scenarios well enough such that it can bootstrap learning and adaptation efficiently. In this talk, I’ll present examples of FAIR robotics research towards the goal of learning general representations for a wide spectrum of robotics applications.

Robotists face many open design challenges, such as complete degradability and biocompatibility, arbitrary shape morphing, and freely reconfigurable degrees of freedom. In this talk, I will present that addressing these challenges sometimes requires looking beyond conventional mechatronic systems. I will present several examples, ranging from purely passive, material-driven field “robots” to material-mechanism hybrid systems with tunable degrees of freedom. Additionally, I will discuss the critical role of computational design and optimization in assisting robotic and machine design and control. Through these examples, I hope to share the perspective that, when designed and engineered strategically, morphing materials and mechanisms can empower machines with embodied intelligence.

In the rapidly advancing realm of autonomous driving, developing and deploying efficient and scalable perception models across vehicle platforms is crucial. This presentation introduces a Unified Perception Model crafted to manage the diverse array of perception tasks essential for autonomous vehicles, such as object detection and occupancy estimation, within a single, integrated framework. While there are many topics around the Unified Perception Model, we will focus on the efficiency and scalability of the model. In particular, you will find answers to the following two questions: 1) How can we leverage temporal information without sacrificing training efficiency and model capacity? 2) How to do large scale cross-platform pretraining and cross-platform deployment? Our unified model aims to streamline operations, reduce complexity, and enhance adaptability across diverse autonomous driving platforms, marking a significant advance in the development of autonomous vehicle technology.