Feb 10 - Physical AI Meetup: Robot Brains and Models
Raleigh AI, Machine Learning and Computer Vision Meetup Join us for the Physical AI, ML and Computer Vision Meetup series focused on robotics use cases. In this session we will focus on robot brains and models.
Date, Time and Location
Feb 10, 2027 9:00 AM - 11:00 AM PST Online. Register for the Zoom!
Physical AI That Adapts
Every vision model is trained on a snapshot of the world, but robots are deployed into one that keeps changing: lighting shifts, friction drifts, objects move, actuators wear. Today the fix is a slow loop of collecting new data, retraining, validating and redeploying, which can take weeks or months. In this talk, we show how Lazy Dynamics’ LazyInfer closes that gap by treating perception, parameter learning and control as one Bayesian inference problem that runs on the device, inside the control loop.
About the Speaker
Bart van Erp is CTO at Lazy Dynamics. He has a background in Electrical Engineering and a PhD in the field of probabilistic machine learning.
Robotic World Models: Cloud Versus On-Edge
Cloud-hosted robotic world models can draw on substantial compute, while on-edge models promise local, low-latency inference; choosing between them depends on the reliability of their predictions. We examine this question through a from-scratch, gradient-verified reproduction of the proprioceptive Robotic World Model for ANYmal D, using recorded robot trajectories and CPU resources. Autoregressive training reproduces the model's long-horizon advantage over teacher forcing, while ensemble disagreement tracks the ordering of prediction errors but substantially underestimates their magnitude. On the released checkpoint's training episodes, disagreement is 33.4 times smaller than realized error at the method's 100-step rollout horizon, making it a useful ranking signal but not a calibrated error bar.
About the Speaker
Joyjeet Singh is an independent researcher in physics-informed machine learning and the sole author of three arXiv preprints on implicit-reasoning models, world-model reproducibility, and the planning objectives used over learned latent dynamics.
Faster Physical AI Development with CoreWeave
Robot brains are getting bigger. Vision language action models, world models and foundation policies all demand large training runs, fast experimentation and serving that keeps up with real robots. In this session, Narendra Kumar Vadapalli, Senior Specialist Field Engineer at CoreWeave, walks through how physical AI teams are building the infrastructure behind these models today. Drawing on his work helping customers design and scale AI platforms on Kubernetes, he will cover how teams organize multimodal training data, run distributed training and simulation side by side, and move models from experiment to deployment without rebuilding their stack each time. Attendees will leave with practical patterns for scaling robot learning workloads and a look at what is coming next.
About the Speaker
Narendra Kumar Vadapalli is a Senior Specialist Field Engineer at CoreWeave, where he has helped customers design and scale AI infrastructure since 2024.
Robot Brains and Models with Voxel51
About the Speaker
Brent Griffin is a Principal Machine Learning Scientist at Voxel51, specializing in low-cost learning on unstructured data at scale. He previously built and led the perception group at Agility Robotics and was an assistant research scientist at the University of Michigan, working at the intersection of computer vision, control, and robot learning.