Physical AI has moved from niche to necessity.
The second day of Actuate, the developer conference for Physical AI, built on the momentum of Day 1 discussing massive automated warehouses to generalist foundation models, every session made it clear: Physical AI has moved from niche to necessity.
James Kuffner, CTO, Symbotic
James Kuffner’s talk on high-scale robotics orchestration highlighted the real-time coordination of thousands of mobile robots across massive logistics networks—and how the convergence of machine learning, perception, and multi-agent planning is redefining warehouse automation.
RJ Scaringe, CEO, Rivian
Adrian Macneil, CEO, Foxglove
In a wide-ranging conversation, RJ Scaringe described Rivian’s founding journey, its vertically integrated strategy, and the path forward for sustainable transport. Scaringe emphasized the role of software-defined vehicles and autonomy in reshaping not just individual cars, but entire mobility ecosystems. From first principles engineering to production at scale, the message was clear: physical products and intelligent systems must now evolve together.
Boris Sofman, CEO, Bedrock Robotics
Boris Sofman shared Bedrock Robotics’ vision for full autonomy in the construction sector—a field where productivity has stagnated for decades. Drawing on his experience at Waymo and Zoox, Sofman illustrated how the team is adapting autonomous driving techniques to rugged, unpredictable, and highly dynamic construction environments.
Bedrock’s approach focuses on retrofitting and modernizing existing heavy equipment with autonomy capabilities, enabling safer, more productive job sites without reinventing the hardware.
Kay Ke & Sergey Levine, Physical Intelligence
Sergey Levine and Kay Ke returned to Actuate to unveil the latest from Physical Intelligence: robotic foundation models designed to operate any robot, in any setting, with minimal fine-tuning.
They shared details on their second-generation Vision-Language-Action (VLA) models, hierarchical planning systems, and large-scale training pipelines. These models aim to provide generalist capabilities to robotic systems, just as LLMs have to language—enabling broad task coverage and robust zero-shot behavior in physical environments.
Rajesh Radhakrishnan, Head of Autonomy, Serve Robotics
Serve Robotics is on track to deploy 2,000 sidewalk delivery robots in 2025. Rajesh described how their autonomy stack evolved to support fast iteration, learning across deployments, and reliable behavior in noisy, GPS-denied urban settings.
Key enablers included a fleet-level data flywheel, deployment-aware simulation environments, and a modular AI architecture that allowed quick adaptation to new cities, traffic laws, and pedestrian behavior patterns.
A panel hosted by Chris Paxton (Agility Robotics) featuring, Evan Morikawa (Generalist), Shitij Kumar (Dexterity), John Betancourt (Kiwibot) wrestled with a core tension in modern robotics: how much structure do we need to build in, versus how much should we learn from data? On one end, VLA models promise end-to-end generalization; on the other, engineered systems offer precision, predictability, and clarity.
Panelists debated abstraction boundaries, safety tradeoffs, and practical realities of deploying systems that must perform reliably in the wild. There was no clear winner—only alignment that hybrid approaches, grounded in data but guided by structure, are likely to prevail in the near term.
Bahram Banisadr, Head of Product at Foxglove
Bahram discussed what’s still missing in robotics infrastructure—and what Foxglove is doing to address it. From streamlining observability to standardizing message formats and reducing integration tax, the talk introduced new features designed to help robotics teams move faster and scale smarter.
Read the full product announcements here.
Mike Brevoort, Principal Engineer – AI Lead, Mytra
Mike Brevoort’s talk explored how real-world robotics deployments rarely follow clean, top-down designs. Drawing from his work at Mytra, where AI is used to coordinate fleets of autonomous mobile robots in dynamic warehouse environments, Mike highlighted how robotic systems evolve through layers of operational hacks, safety adaptations, and software abstractions that accumulate over time. Rather than fight this complexity, Mytra embraces it, treating robotic systems as living organisms shaped by their constraints, constantly adapting through feedback loops, edge-case handling, and on-the-fly orchestration.
Mahesh Krishnamurthi, CPO, Vayu Robotics
Deploying end-to-end models in the real world requires more than accuracy—it requires safety, transparency, and interpretability. Mahesh introduced Vayu’s three-part strategy:
Together, these systems help Vayu deploy foundation models in public environments while maintaining high standards of operational safety.
Robert Sun, Founding Engineer, Dexterity
Robert made the case that true intelligence in robotics demands rethinking everything—not just the AI model, but its embodiment and the world it operates in.
He introduced IRIS, a unified architecture that tightly couples reasoning, hardware control, and environmental adaptation. Rather than retrofit AI to existing machines, IRIS builds the stack from the ground up to maximize competence in real work—whether that means palletizing fragile goods or performing nuanced warehouse tasks.
This philosophy reframes Physical AI: not as an emulation of human ability, but as a new category of skilled, purpose-built systems optimized for the jobs we actually need done.
Deepak Pathak, CEO & Co-Founder, Skild AI
Deepak closed the technical track with an ambitious vision: one foundation model that can control any robot, on any task.
Backed by his work in self-supervised learning and sim2real transfer, Skild AI is developing general-purpose control models trained on broad multi-task datasets. Pathak outlined their architectural innovations, training strategies, and results from early deployments, showing that general robotic intelligence is not just theoretical—it’s already being tested in the field.
To close out the conference, Rivian hosted a reception at its San Francisco showroom, bringing together founders, engineers, researchers, and operators for one last round of networking and toasts to the future of robotics.
If Day 1 was about building the pieces, Day 2 was about putting them together. From scalable autonomy stacks and robotic foundation models to real-world orchestration at industrial scale, Actuate 2025 showed that robotics has crossed the threshold: it’s no longer about if—it’s about how fast we can scale.
Physical AI isn’t a research problem anymore. It’s a systems problem. And at Actuate, we saw the people solving it.
Subscribe to Foxglove’s newsletter to stay up to date on future product and Actuate announcements.
