Robot safety is a nuanced concept. We commonly equate safety with collision-avoidance, but in complex, real-world environments (i.e., the “open world’’) it can be much more: for example, a mobile manipulator should understand when it is not confident about a requested task, that areas roped off by caution tape should never be breached, and that objects should be gently pulled from clutter to prevent falling. However, designing robots that have such a nuanced safety understanding---and can reliably generate appropriate actions---is an outstanding challenge. In this talk, I will describe my group’s work on systematically uniting modern machine learning models (such as large vision-language models and latent world models) with classical formulations of safety in the control literature to generalize safe robot decision-making to increasingly open world interactions. Throughout the talk, I will present experimental instantiations of these ideas in domains like vision-based navigation and robotic manipulation.

Our work is broadly motivated by the emergence of learning-based methods in control theory and robotics, with a specific focus on scenarios that have humans in-the-loop with control systems. For instance, learning algorithms are being deployed in semi-autonomous vehicles, robot assistants, brain-machine interfaces, and exoskeletons, where they interact dynamically with a human partner to complete tasks. When learning algorithms are employed in this way, a dynamic game is created that is played between two intelligent agents (the human and machine learners), requiring new techniques to guarantee safety and performance.

General-purpose robot policies hold immense promise, yet they often struggle to generalize to novel scenarios, particularly struggling with grounding language in the physical world. In this talk, I will first propose a systematic taxonomy of robot generalization, providing a framework for understanding and evaluating current state-of-the-art generalist policies. This taxonomy highlights key limitations and areas for improvement. I will then discuss a simple idea for improving the steerability of these policies by improving language grounding in robotic manipulation and navigation. Finally, I will present our recent effort in applying these principles to scaling up generalist policy learning for dexterous manipulation.

Many of us are collecting large scale multitask teleop demonstration data for manipulation, with the belief that it can enable rapidly deploying robots in novel applications and delivering robustness in the 'open world'. But rigorous evaluation of these models is a bottleneck. In this talk, I'll describe our recent efforts at TRI to quantify some of the key 'multitask hypotheses', and some of the tools that we've built in order to make key decisions about data, architecture, and hyperparameters more quickly and with more confidence. And, of course, I’ll bring some cool robot videos.

Aerial robotics have become ubiquitous, but (like most robots) they still struggle to operate at high speed in unstructured, cramped environments. By considering a vehicle's mechanical design simultaneously with the design of controls and automation algorithms, we have more degrees of freedom to find creative solutions to problems. In this talk I will present some of my group's work on enhancing aerial robots, including purely algorithmic approaches ('how can I do more with the hardware I already have?') and with hardware co-design ('how can I change the vehicle so that the hard problem is actually easy?'). I will discuss two exemplary challenges for aerial robots: first: flight through narrow, unstructured environments, and second: long duration and range flight within the constraints of battery-electric power. Lastly, I will discuss some work on adaptive and learning control, specifically for robustness to parametric uncertainty. For flight through narrow environments, I will present an algorithmic approach for high speed path planning that incorporates perception uncertainty, and can be used on a standard drone. We will then present two alternative approaches that modify the system design: one a vehicle that can change its shape to fit through narrower spaces, and a second that is highly collision resilient, and for whom collisions are therefore neither mission- nor safety-critical. For overcoming energetic challenges, we will present a strategy for real-time optimization of flight characteristics for a vehicle, specifically using extremum seeking control to modfiy the system airspeed and yaw angle; an algorithm that can be applied to any aerial robot. We then again show two design modifications to work around the problem -- first, a morphing system that can reduce its drag area at speed, and secondly a system capable of mid-air battery replacement for indefinite flight.

Controlling large teams of robots is a crucial challenge in robotics, due to the need for solutions that balance efficiency, scalability, and robustness. This talk will delve into recent advancements in learning-based control for multi-robot systems, with a focus on scalable coordination and decentralized decision-making. I will present the latest results we have achieved to efficiently learn distributed control strategies for large teams of robots, leveraging physics-informed machine learning and generative AI techniques. We will see how physics-informed learning can be used to provide the learned controllers three key properties: interpretability, modularity, and scalability. Similarly, we will demonstrate how generative AI can be used to ease interaction with non-expert users to describe desired large-scale swarm configurations, producing smooth trajectories and accounting for potential collisions through a reactive navigation algorithm.

Intelligence does not emerge fully formed, but it forms from a developmental cycle. From the earliest stages of life, animals acquire intelligence through cycles of sensing, adapting, and connecting, with their bodies, their environments, and each other. This embodied developmental process is not just a feature of natural intelligence; it offers a powerful blueprint for designing machines that are more adaptive, generalizable, and capable of meaningful interaction in the real world. In this talk, I will present a developmental arc of embodied intelligence centered on three interdependent capacities: Sense, Adapt, and Connect. I will begin with interactive perception systems that uses sound, vibrations, touch, vision, and smell to construct a comprehensive understanding of the environment. I will then move to adaptation, arguing that robust generalization in robots require modeling the “self”, both behaviorally and physically. Such understandings of the self will enable robots to reflect on its behavior, understand its morphology, and adjust in response to change. Finally, I will briefly discuss how connection extends intelligence into the social domain, where machines must synchronize, communicate, and collaborate with others in a shared world. Together, these ideas represent a unified vision: to build machines that grow, rather than are assembled — machines that do not just function, but evolve, recover, and relate. I will close by discussing how this developmental perspective may inform the future of intelligent machines.

Building robots that can operate autonomously in unstructured environments by following arbitrary natural language commands has long been the north star in robotic manipulation. While there has been tremendous progress in learning visuomotor policies that exhibit promising signs for open-world deployment, generalization to unseen tasks or motions largely remains unattainable or out of scope. In this talk, I will discuss how deliberate choices of task representations enable such zero-shot generalization at the task level, despite given no task-specific demonstrations. Notably, I will discuss our years-long investigations into extracting task representations from off-the-shelf foundation models; I will discuss its evolution from a language-only representation to 4D space-time domain and their applications to model-based planning, affordance learning, and visuomotor policy learning. At the end of the talk, I will present an alternative view for scaling towards robotic intelligence: by leveraging foundation models to provide task-specific knowledge in the form of task representations, robotic data scaling can focus on learning from task-agnostic interactions with a world modeling objective, such that collectively this enables robots that not only understand the world as humans do but can also act within it with purpose and generality.

Recent advances in robot learning have mirrored the progress of large language models in many ways—yet, one key distinction remains: action chunking. In this talk, I will begin with an analysis of action chunking, highlighting its inherent tradeoff between long-term consistency and short-term reactivity. I will then introduce two methods to address this tradeoff: (i) Bidirectional Decoding: an inference algorithm that jointly optimizes consistency and reactivity using additional compute at test time; (ii) Past-Token Prediction: an auxiliary training objective that encourages diffusion policies to capture temporal dependencies in long-context observations. Together, these methods offer a promising path toward memory-aware robot policies without action chunking.

Soft robots locomote by reshaping compliant bodies. However, precision control of large, reversible 3-D deformations with minimal hardware remains difficult. We tackle this by treating morphing as a geometric problem: prescribing an in-plane “metric actuation” field—coordinated local area expansions and contractions—changes a surface’s intrinsic curvature, causing a flat surface to autonomously adopt a prescribed 3-D form. For instance, a circular elastomer disk whose center undergoes a four-fold isotropic areal stretch while its rim stays fixed reliably morphs into a hemisphere; the same curvature distortion, applied at any scale, yields identical shapes because only relative strain matters. We convert desired shapes into these metric maps and demonstrate them in two soft-robotic platforms: (i) microscale Polyimide layer patterned using semiconductor lithography that transform into free-standing doubly-curved electronics, and (ii) cm-scale silicone surfaces that are laser-cut and deploy from flat packs into load-bearing domes suitable for extraterrestrial shelters. Because the curvature “code” is embedded in the 2-D layout, actuation collapses to a single global stimulus - pressure, temperature, or gravity - dramatically simplifying control schemes and opening clear paths to soft grippers, morphing airfoils, and adaptive wearables.