Ray Summit 2026 Recap: Scaling Reinforcement Learning at Scale

Yesterday, we announced Antioch's $32 million Series A to help teams build physical AI systems at the speed of software. We’re pursuing that goal by combining the control and coverage of classical simulation with models that learn from real-world data. Classical simulators can’t capture every physical interaction, while fully learned models are often data-constrained. Antioch's approach combines the best of both worlds: a hybrid simulation that uses available data to learn the behaviors that are hardest to model explicitly. As those simulations become more accurate, teams can develop and deploy faster, generating new real-world data that feeds back into the simulator. Over time, this loop helps us learn more complex interactions and train larger parts of the simulator together, moving us toward end-to-end learned world models. The work that helps teams ship today also helps us build the next generation of simulation. We’re working with Amazon Ring, NVIDIA Robotics, and Nebius to put this into practice. Take a read through our raise announcement blog post to learn more about our approach and where we’re headed. Link in the first comment.

Can embedding physical AI engineers directly into your R&D lab turn messy sensor data and live telemetry into game-changing operational speed? Following CoreWeave's rollout of Physical AI Field Engineering, domain experts are working on site with teams like the Aston Martin F1 Team and Nissan Motor Corporation to deploy validated, physics-backed models. Discover how Richard Ahlfeld, Ph.D. (CoreWeave), Emma Deutsch (Nissan Motor Corporation) and NVIDIA CEO Jensen Huang view the shift toward general-purpose robotics, agentic learning, and real-time telemetry processing in high-performance environments. Read more: https://lnkd.in/eRf_MQVE #ArtificialIntelligence #TechTrends #CoreWeave #AstonMartinF1

Isaac ROS 5.0 was launched today at #ROSCon in Toronto. 🚀 Noble Machines is using #NVIDIA Isaac ROS on #Jetson to accelerate the development of our general-purpose robot, Moby — adopting ready-to-use AI and perception capabilities rather than creating them from scratch. From our CEO Wei Ding: "Putting general-purpose robots to industrial work comes down to two frontiers that we have to push: high-performance hardware and the AI that drives it. Isaac ROS gives us tools we can adopt rather than reinvent, accelerating our ability to build those competitive advantages — including GPU-accelerated perception, TensorRT inference, and hardware codecs, all native to the Jetson edge computing platform. Isaac ROS accelerates us, and in turn we can better accelerate the #physicalAI ecosystem.” For more on how Noble Machines is building a learning stack for industrial autonomy, read the team's blog ➡️ https://lnkd.in/gEr7VUWn Isaac ROS 5.0 launch blog 🔗 : https://nvda.ws/4cWwNkV #Robotics #Manufacturing NVIDIA Robotics

Nessco's Physical AI framework is tackling robotics training head-on with a 100-robot grid designed specifically for high-throughput AI training. www.nessco.ai By integrating a "hyperbolic time chamber" simulation environment, engineering teams can train and iterate through tens of thousands of robot models at unprecedented speeds. Scaling Physical AI from the lab to production requires infrastructure built for massive parallel simulations: 1️⃣100-Robot Grid Architecture: Concurrent multi-agent training pipelines that drastically accelerate reinforcement learning and policy convergence. 2️⃣Hyperbolic Time Chamber Training: Compressing months of physical trial-and-error into accelerated virtual cycles, allowing teams to evaluate thousands of model variants concurrently. 3️⃣Isaac Sim Compatibility: Native compatibility with NVIDIA Isaac Sim and Omniverse ecosystems, ensuring high-fidelity physics, accurate sensor simulation, and seamless synthetic data generation. The future of autonomous systems relies on closing the sim-to-real gap faster than ever before. How is your team tackling simulation scaling and synthetic data pipelines? #PhysicalAI #Robotics #NVIDIAIsaacSim #EmbodiedAI #ArtificialIntelligence #MachineLearning #TechInnovation

With the iconic TECH WEEK by a16z around the corner, we couldn't think of a better time to launch Industrial Next (YC W22)'s Real2Sim2Real (R2S2R) Arena, a Physical AI development and testing facility at our 10,000 sq ft San Francisco HQ. With more than 50 semi-humanoids, the Arena brings simulation, robot learning, and real-world deployment under one roof for teams at the frontier of Physical AI. Physical AI remains divided between simulation and the real world, where contact dynamics, object variation, hardware differences, and unexpected conditions affect performance. The Arena brings these stages together in a continuous loop: every real-world task gets recreated in simulation, simulation scales training and validation, and the final gap gets closed via real-world deployment in the arena. We're launching with support from partners including NVIDIA Robotics, whose Isaac Sim platform and the Newton physics engine anchor much of our simulation stack, along with Versor, ALLSIDES, Paddy, the University of Washington’s Personal Robotics Lab, and the University of Cambridge’s CamRAL (the Cambridge Resilient Autonomous Learning Lab), with more to be announced soon. Public access will open on launch day through sim2world, a benchmarking initiative designed to measure how well policies transfer from simulation to the real world. Researchers and companies will be able to contribute tasks, train in simulation, validate on physical robots, and help shape the Real2Sim2Real benchmark, with a portion of capacity reserved for public research. 📅 Thursday, October 8, 2026 🕓 4:00–7:00 PM 📍Industrial Next - San Francisco 🎟️ RSVP - Limited capacity: https://lnkd.in/gKD3YYQ9 Want to contribute a task, conduct research, or partner with the Arena? Shoot me a message 📬 #R2S2R #PhysicalAI #Robotics #Manufacturing Andreessen Horowitz | Gilwoo | Lukas | Allen

Is your company considering a discovery project around humanoid robots? If so I would be happy to discuss the hype, the promises and the reality of the industry as it stands today. We are moving our robots out of research labs and into warehouses to experiment with scanning, picking, moving and packing products.

We Teleoperate Until We Reach 100% Automation Yuri starts working the day it arrives, teleoperated by a person doing the task through the robot. You take the controls and do the task through Yuri, reaching across the table, picking the item up, carrying it over, setting it down. Sensori’s Telepath software captures every task and motion of the robot as training data to build Physical AI. Yuri’s onboard NVIDIA AGX computer brain allows it to do tasks autonomously while a tele-operator stands by to correct any mistakes or hiccups. Even the corrective actions of the operator trains Yuri how to do the job better next time. Download a free white paper titled: Evaluating a Research Robot for Physical AI here: https://loom.ly/tVUprLo. Designed, assembled, shipped and supported in Southlake, Texas. NVIDIA Robotics #Robotics #PhysicalAI #EmbodiedAI #MobileManipulation

Physical AI, the next frontier of Intelligence, in real Engineering world ( Touches multiple areas from Robotics to Autonomus Vehicles, from CAE Simulations to AI Led Simulations for stress/life predictions and digital twin/digital thread for complete life cycle management of products). Important role that Vision AI and Edge Computing, IT/OT Integration for Industry4.0 plays in physical AI world, is touched upon. Physical AI brings Gen AI and Agentic Intelligence into physical world, enabling machines to sense, think and react. This course helps, building skill to design and deploy intelligent physical systems, across enterprise.

NVIDIA #Industry #Safety Hi, the Industrial AI Podcast is back from summer holidays. Robert talks to Dr. Riccardo Mariani about AI and safety in the industrial world. Riccardo explains how NVIDIA is shaping new standards for physical AI and why safety is not just a requirement, but a value driver in modern automation. And he also shares insights on the importance of end-to-end safety frameworks, the role of simulation, and the integration of AI into robotics and machinery.