Locomotion in complex environments (e.g., rubble, leaf litter, granular media) is essential to mobile engineered systems such as robots. Effective locomotion requires complex control strategies to interact with terrain heterogeneity. Computational intelligence (CI), which typically includes rapid terrain sensing and active feedback controls, is a widely recognized component in locomotion strategy. Alternatively, mechanical intelligence (MI) - passive response to environmental perturbation governed by physics laws or mechanical constraints - is an important yet less studied component. In this talk, I will discuss 'why' and 'how' MI can contribute to effective locomotion using the examples of multi-legged robots (redundantly segmented bodies with simple legs). For the 'why,' I will quantify a specific MI that emerges from leg redundancy. By modeling locomotion as a stochastic process (analogous to signal transmission over noisy channels), I will show that MI, without any CI, is sufficient to generate reliable and effective locomotion. To explore the 'how,' I will take a quantitative analogy to signal transmission algorithms (e.g., error correcting/detecting codes) and propose a co-design coding scheme for multi-legged locomotion. Specifically, my talk will cover that (i) additional legs, with higher control dimensions, can enable a broader spectrum of capabilities, including load carrying/pulling, sidewinding, rolling, and obstacle-climbing; (ii) the inclusion of CI (feedback controls) can enhance multi-legged locomotion speed while preserving the feature of robustness; and (iii) CI might reduce the number of redundant legs required to navigate a particular terrain. Finally, I will discuss the coordination and competition between MI and CI in a broader framework termed Embedded Intelligence (EI) and illustrate the applications of MI-dominated systems in fields like search-and-rescue, agriculture, and the development of soft, micro, and modular robots.

Modern AI advances by transferring knowledge from humans to machines at scale. Vision and language models learn from vast Internet data, but robot learning still relies heavily on slow, labor-intensive teleoperation. Recently this assumption has begun to shift: growing industrial efforts are collecting large amounts of human experience data to scale robot performance. As large-scale data collection becomes increasingly feasible, the central challenge shifts to understanding how robots can learn from human behavior. In this talk, I argue that human-to-robot transfer can be understood as two coupled problems: extracting priors about physical intelligence from human experience, and grounding those priors into a robot’s embodiment. I will revisit several of our recent works through this lens, showing how egocentric human data enables scalable learning of manipulation priors, while representation learning and cross-embodiment transfer address the grounding challenge. I will also discuss recent results showing emergent human-to-robot transfer from large-scale human pretraining, as well as evidence that learning across diverse robot embodiments can further improve transfer. Finally, I will introduce EgoVerse, an ecosystem for robot learning from embodied human data, and discuss how collaborative platforms can enable both rigorous science and organic data growth. I will conclude with future directions toward more human-centered robots that better understand human intent and collaborate naturally with people.

Today's robots can perform impressive short-horizon tasks, but they're still far from the kind of autonomous agents we've now grown used to in the digital world; systems you can hand a complex, multi-step job and trust to figure it out. In this talk, I'll argue that getting there requires two ingredients we've historically struggled with: giving robot policies a sense of memory, and training generalist behaviors that are both broadly capable and high-performing. I'll discuss our recent progress on both fronts through two works, π0.6-MEM and π0.7, and sketch what's still missing on the path to long-horizon physical autonomy. Bio: Karl Pertsch is a member of the technical staff at Physical Intelligence. Before, he was a postdoc at UC Berkeley and Stanford. His work focuses on building generalist robot policies that can solve a wide range of physical manipulation tasks in the real world. Karl obtained his PhD from USC. During his PhD he interned at MetaAI and Google Brain. His work has been awarded the Best Conference Paper Award at ICRA'24, two Outstanding Paper Awards Finalists at CoRL'24, and a Best Paper Finalist at RSS'25.

Despite remarkable progress in clinical adoption over the past two decades, surgical robotic systems remain large, physically constrained, and limited in autonomy. Today's commercial platforms are teleoperated and mechanically restricted to specific procedures, requiring trained teams to set up and operate them, and offering limited medical image integration and AI guidance. These conventional systems are also expensive and challenging to deploy and difficult to scale. Meeting the evolving needs of modern healthcare requires rethinking not just how robots are used in surgery, but what they can become—shifting from rigid, teleoperated tools toward fully autonomous robot partners in the operating room. This talk presents our research on enabling greater surgical autonomy through the lense of robot learning. I will close with our recent work on human-robot teaming using humanoid systems, examining the limitations of the current state of the art and the opportunities ahead for research and deployment.

To truly transform our lives, autonomous systems must operate in complex environments shared with other agents. For instance, delivery robots need to move through spaces that are shared with humans, and warehouse robots must coordinate in shared factory floors. Such multi-agent settings demand systematic methods that enable efficient and reliable interactions among agents. In the first part of my talk, I focus on control challenges in such domains and will discuss game-theoretic planning and control for robots. Intelligent interaction requires robots to reason about how their decisions affect and are affected by others. I will present our recent results showing how exploiting structural properties of interactions leads to motion planning algorithms that are both efficient and tractbale for real-time deployment. The second part of the talk will focus on learning in interactive domains, including imitation learning and reinforcement learning. While these approaches have advanced significantly in single-agent settings, multi-agent domains present unique learning challenges because decisions are tightly coupled across agents. I will highlight how some of these challenges can be addressed to make learning feasible in interactive multi-agent domains.

I aim to build complete intelligent agents that can continually learn, reason, and plan: answer queries, infer human intentions, and make long-horizon plans spanning hours to days. In this talk, I will describe a general learning and reasoning framework based on neuro-symbolic concepts. Drawing inspiration from theories and studies in cognitive science, neuro-symbolic concepts serve as compositional abstractions of the physical world, representing object properties, relations, and affordances, and actions. These concepts can be combinatorially reused in flexible and novel ways. Technically, each neuro-symbolic concept is represented as a combination of symbolic programs, which define how concepts can be structurally combined (similar to the ways that words form sentences in human language), and modular neural networks, which ground concept names in sensory inputs and agent actions. I show that systems that leverage neuro-symbolic concepts demonstrate superior data efficiency, enable agents to reason and plan more quickly, and achieve strong generalization in novel situations and for novel goals.

While learning algorithms have powered tremendous progress in autonomy in recent years, failures still persist when deployed systems inevitably encounter rare scenarios dissimilar from training data—so-called out-of-distribution (OOD) scenarios. We will present a holistic perspective on the use of guardrails to mitigate the negative consequences of such OOD scenarios on the safety and reliability of learning-based robotic systems, organized around three questions: how to detect when a system is unreliable, how to enact safety interventions when it degrades, and how to improve it through principled use of deployment experience. First, to detect impending failures, we show how the commonsense reasoning capabilities of foundation models, e.g., large language models (LLMs), make them attractive as runtime monitors for intricate, context-driven hazards. Second, we introduce a 'Thinking, Fast & Slow' reasoning framework that tightly integrates LLMs into the robotʼs control loop, thereby enabling their reactive, real-time use to avert closed-loop failures caused by system-level deficiencies in the robot's semantic reasoning. Finally, beyond averting immediate failures during deployment, we aim to address observed failure modes at their source by training improved policies. Towards this goal, we trace policy successes and failures back to individual training samples, enabling data level diagnosis and data curation that leads to improved policy performance and robustness. Taken together, these contributions advance towards the goal of trustworthy autonomous systems: systems that know what they can and cannot do, recover safely when conditions become too challenging, and can be interpretably analyzed and improved when failure modes emerge.

Human are naturally dexterous, and human behavior provides a rich source of data for robot learning. Even for a simple task such as rotating a ball, there are many possible strategies: using finger motions, rotating the wrist, or performing repeated regrasps with a parallel gripper. The human-like strategy is often the most desirable not because it is the only solution, but because humans provide one of the most abundant and expressive sources of demonstrations for learning general-purpose manipulation. To effectively leverage human data, robots must develop dexterity that can align with human demonstrations while accounting for their own embodiment. In this talk, I will discuss two directions toward this goal: 1) learning dexterous teleoperation from humans, and 2) learning dexterous skills from human videos. I will show how these approaches can bridge human motion and robotic control, enabling robots to acquire more human-like dexterous skills.

oday, most robot learning models ignore the geometry of the environment, focusing instead on enlarging model and dataset size. Whereas this approach has been very successful in language and vision, geometry is more important in robotics than in those fields. How can we leverage geometric sructural priors in robot learning? In this talk, I will summarize our recent work on this topic and make predictions about the future.

Robotics is advancing rapidly, yet today's robots still struggle to match the dexterity of humans. A promising path toward autonomous dexterous robots is through learning-based methods trained on data collected via teleoperation, where a human operator controls a robot in real time to demonstrate complex tasks. In this talk, I will present the current state of dexterous teleoperation research and propose the field’s biggest challenges and opportunities.