Austin Robotics & AI

Hello Everyone,

There are many exciting developments in Robotics and AI right now, so let's get together to learn, connect, and create more opportunities for everyone here in Austin.

This event will be held in-person at HICAM from 6 pm - 8 pm.

Sign up here: https://www.hicam.io/event/austin-robotics-and-ai-february

Speaker 1: Bhairavsingh Ghorpade, Senior Data Scientist at Space Cow LLC

Talk 1: Enhancing BRepMFR for Machining Feature Recognition via Vision Transformers and Geometry-Aware Fourier Positional Encoding

Summary (Abstract): Accurate interpretation of geometry in 3D Boundary Representation (B-rep) CAD models is essential for AI-based machining feature recognition and process planning. The BRepMFR deep learning architecture demonstrated that transforming CAD models to face-adjacency graphs and applying a Graph Transformer is an effective approach to learning geometric and topological representations supporting machining feature recognition. However, this architecture relies on convolutional encoders and handcrafted descriptors that struggle to capture the global curvature and geometric variability inherent in industrial CAD models. To address these limitations, we introduce a Vision Transformer (ViT)–based surface encoder with geometry-aware Fourier positional encoding and Laplacian Positional Encoding (LapPE). The ViT replaces the original CNN-based encoder, enabling global attention over UV-parametric patches and producing richer surface embeddings. To better model long-range geometric correlations and topological relationships, Fourier-encoded pairwise distances and graph Laplacian eigenvectors augment node and edge features.

We trained on 10,000 synthetic CAD models from the CADSynth dataset, covering a wide variety of machining-relevant geometric configurations. Experimental results showed improved feature classification accuracy, stronger class separation, and more stable attention patterns across complex surfaces such as cylindrical features. Gains from the original BRepMFR are observed in various classes such as the Triangular Passage class, which improved from an F1 of 0.855 to 0.903 due to stronger recall.

Space Cow LLC is an applied research company with skills in software, data, and AI along with domain expertise in manufacturing, industrials, and defense. We seek to develop innovative capabilities and solutions with emerging technologies.

--------

Speaker 2: Dr. Nathan Tsoi - Assistant Professor of Practice and Postdoctoral Fellow at Texas Robotics at the University of Texas at Austin

Talk 2: From Tools to Teammates: Aligning Robot Learning with Human Perception

Summary (Abstract) As mobile robots move beyond the lab and into public spaces, they must do more than just avoid collisions; they must behave in a socially acceptable manner. However, current robot learning approaches often treat value alignment as a unidirectional problem, where the agent passively learns a fixed objective from a human. This talk argues for a new perspective: Bidirectional Value Alignment, which models interaction as a continuous loop where the robot’s actions actively shape the human’s internal state, including elements such as trust, comfort, and expectations.

I will discuss how we can move beyond standard measures to train systems and optimize for social success. We will explore a formulation where the human is treated as a dynamic partner with latent states that the robot must estimate and manage. By incorporating these human elements directly into the robot's decision-making process, we can transition from building obedient tools to designing socially aware teammates that foster long-term, high-trust partnerships.

Speaker Bio:
Dr. Nathan Tsoi is an Assistant Professor of Practice and Postdoctoral Fellow at Texas Robotics at the University of Texas at Austin, advised by Joydeep Biswas and Peter Stone. His research focuses on enabling the widespread deployment of mobile robots that are capable of completing tasks reliably while interacting naturally with people. He develops robot learning systems that are sample efficient, robust to real-world uncertainty, and grounded in methods that optimize for human-centric values such as competence, comfort, and social appropriateness.

Nathan received his PhD in Computer Science from Yale University, advised by Marynel Vázquez, where his work introduced new simulation platforms and data-driven methods for aligning robot behavior with human social values. Previously, he worked as a research engineer at the Stanford AI Lab and as a senior software engineer at Sequoia Capital.

Sign up here: https://www.hicam.io/event/austin-robotics-and-ai-february