Stanford Robotics Seminar ENGR319 | Spring 2026 | Leveraging Geometry in Robot Learning

Why It Matters

Integrating geometric priors into learning models promises faster, more reliable robot deployment while cutting the data and compute costs that currently limit large‑scale vision‑language systems.

Key Takeaways

• Geometry‑aware models bridge hand‑coded and data‑hungry robot learning.
• Equivariant layers embed physical symmetries, reducing required training data.
• Point‑cloud, spherical, and ray representations improve policy learning efficiency.
• Geometric transformer attention incorporates reference frames into action prediction.
• Hybrid approaches achieve one‑shot manipulation with fewer demonstrations.

Summary

The seminar examined the growing divide between traditional hand‑coded geometric models and modern vision‑language models (VLMs) in robotics. While classic approaches rely on precise, physics‑based priors that enable one‑shot tasks, they falter when reality deviates from assumptions. Conversely, today’s VLMs learn directly from massive datasets but discard explicit geometry, demanding extensive training to recover spatial reasoning.

Rob Platt argued that a middle ground is possible by embedding geometric structure into learning architectures. He highlighted four recent papers that explore point‑cloud encodings, spherical embeddings, 3D‑ray representations, and geometric transformer attention. Central to these methods are equivariant neural‑network layers that respect translation and rotation symmetries, drawing on Emmy Noether’s principle that physical conservation laws correspond to symmetry groups.

Examples included the “Equivariant Diffusion Policy,” which constrains convolutional kernels to a handful of parameters to enforce rotation equivariance, and a spherical‑based policy that operates on SO(3) transformations. The talk also referenced the 2022 RSS best‑paper “Yodo,” demonstrating how a single CAD model can enable one‑shot manipulation when combined with strong geometric priors.

The implication is clear: hybrid models can retain the data efficiency of model‑based planning while leveraging the adaptability of deep learning. By re‑introducing geometry, robots may achieve robust manipulation with far fewer demonstrations, accelerating deployment in industry and reducing the computational burden of training massive VLMs.

Original Description

For more information about Stanford’s graduate programs, visit: https://online.stanford.edu/graduate-education

May 22, 2026

Model Matters: Leveraging Geometry in Robot Learning

This seminar covers:

• How to leverage geometric structural priors in robot learning

• A summary of recent work on this topic

• Predictions about the future

Follow along with the seminar schedule: https://stanfordasl.github.io/robotics_seminar/

Robert Platt is an associate professor in the Khoury College of Computer Sciences at Northeastern University.