General-purpose models trained on large and diverse datasets can often outperform more specialized domain-specific systems, and have demonstrated remarkable capabilities in domains from natural language processing to computer vision. But what would such a generalist model look like in robotics? In this presentation, I will discuss steps toward building broadly applicable robotic foundation models, from assembling suitable datasets, to devising viable and general training objectives, to model architectures that can control a wide range of different robotic platforms, as well as the different ways such models can be used.

At Electric Sheep, we are scaling physical agents across the country to care for our parks and community spaces. Our robots perform tasks such as; mowing, string trimming, and weed treatment using a both foundational world model, ES-1, and Reinforcement Learning. ES-1 is a learned model that consumes time-series information and is capable of predicting both raw features and interpretable outputs useful for outdoor work; such as semantics, a birds-eye-view map, robot pose and traversability. Given this world representation, we can then extract embeddings to teach our agents how to act with RL. In this talk, I will detail how we leverage sim2real techniques to train our models and achieve surprisingly robust simulation transfer across thousands of real yards. I will then delve into how we can leverage language based feedback to improve performance in the real world. 

ROS 2 is a middleware based on a strongly-typed, anonymous publish/subscribe mechanism that allows for message passing between different processes. Since its inception, ROS 2 has used the industry standard DDS as the way to deliver messages both on the same computer, and on the wire to different computers. While DDS has worked for our users, it can be complex to properly configure and understand, and it has a number of shortcomings. To simplify deployment and debugging, the ROS 2 core team has spent a good part of the last year integrating Zenoh into ROS 2. Zenoh is a newer protocol built by some long-time DDS developers, which aims to simplify deployment of networking systems, both locally and across the larger internet. The integration of Zenoh shows great promise as a simpler way for users to configure and use ROS 2.

We will review the data, machine learning, and deployment pipeline necessary to successfully train and validate the models powering Phoenix. As we navigate the uncharted territory of training humanoids, we will share early success stories and hard lessons learned, along with our thoughts on what could accelerate our path to success.

Almost all modern robots and autonomous vehicles rely on sophisticated multimodal sensor arrays. Yet calibrating these complex arrays can be onerous or sometimes impossible without the right approach. For perception engineers, striking a balance among maintaining optimal sensor performance, minimizing operational downtime, and preserving onboard resources has become trickier than ever. In this talk, we explore different technical aspects of both "offline" and "online" (or fully autonomous) calibration configurations, and evaluate how each of these provide different technical, performance and operational advantages (or disadvantages) on both the production line and in the field.

Many modern robotics systems consist of multiple compute devices coordinating to deliver value. At Blue River Technology, the See & Spray system uses 10 synchronized computers with 36 cameras to identify and selectively spray weeds in 10’s of milliseconds. In order to develop & debug such a product, careful and effective data logging is critical. This talk will discuss how we make use of the AOS middleware to produce coherent logs that are robust to failure modes such as intermittent network connectivity or drifting clocks.

Steve Macenski, Open Navigation owner and lead of Nav2, will survey Nav2's techniques and features that set it apart as the world's most deployed mobile robotics autonomy framework. He will then discuss some of Nav2's prominent users and how they have utilized Nav2 in their products and services to rapidly position themselves as industry leaders in their respective domains across logistics, urban environments, agriculture, and much more.

Developing a cutting-edge product feature is a collaborative effort involving multiple stakeholders from initial conception to final release. In this presentation, we will explore the complete journey of creating the "Return to Home" feature for Scythe M.52, our fully autonomous commercial mower. We will discuss the entire lifecycle, starting with initial ideas inspired by vital customer feedback, moving through requirements definition, technical design, testing, and validation. The talk will emphasize the decision-making processes at each stage, the significance of cross-functional collaboration, and the iterative refinements guided by continuous customer feedback. The key takeaway intended from the talk is the translation of robotics into user facing benefits, emphasizing the importance of domain understanding to aid technical decision making.

Labor shortage is a major challenge for the cultivation of work intensive crops. At farm-ng we are developing a camera-based autonomy stack with the help of deep neural networks which aims to enable a variety of autonomous farming operations like scouting, weeding, and harvesting. The talk will give a hands-on overview of our stack and how we use a digital twin and visualization tools to accelerate our development cycle time.

Oceans are hostile, unpredictable environments that present infinite challenges when developing drones capable of autonomously navigating these open waters. This talk from Saronic – a leading developer and manufacturer of autonomous surface vessels for the U.S. Navy – will cover what makes it challenging to build software for seafaring robots, and how they are overcoming those obstacles with the help of simulation, data visualization, and MCAPs.

Collecting, extracting, and analyzing data is a key component of improving robotics systems.  But how do you actually pull data from robots in bandwidth constrained remote environments? And how do you enable your users to analyze that data?  In this talk we will cover our data collection systems on our robots, the data ingestion pipelines to our cloud data lake including how we transform and merge together many different data streams, and the mechanisms we provide to our teams for individual robot debugging via Foxglove, fleet level analysis via SuperSet, product usage in Agility Arc and to support machine learning development for our robotics engineers.

The road to embodiedAI is being paved with impressive end2end trained models. This is particularly remarkable with self driving where we are now transitioning the technology to robust and safe deployment. But what is needed to further accelerate the learning efficiency and universal deployment of end2end models? I will showcase Wayve’s recent research progress to address this broad question.

Brad Porter, CEO and Founder of Collaborative Robotics, in conversation with Foxglove CEO Adrian Macneil, will provide a comprehensive exploration of the critical factors necessary for deploying robotics at scale. Drawing from Brad’s extensive experience leading Amazon’s global robotics operations, they will discuss the considerations and strategies that ensure reliability and scalability in complex environments. Brad will highlight key lessons learned in optimizing operational efficiency and the importance of robust infrastructure to support large-scale robotic systems. The session will also delve into innovative applications that drive industrial automation and provide a forward-looking perspective on emerging trends in the field. Attendees will gain actionable insights into overcoming common scaling challenges and how to apply these strategies to their own robotics deployments.