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

Incorporating geometric understanding—such as physical symmetries and coordinate awareness—into robot learning models makes them significantly more data-efficient and better at generalizing compared to pure "black box" machine learning models.

• Geometric priors — Encoding physical symmetries, such as translation and rotation, allows models to learn from fewer examples by automatically generalizing across different object orientations [12:14].

• Data efficiency — Systems using these structured approaches achieved approximately 10 times better data efficiency than standard baselines, succeeding with fewer than 100 demonstrations [24:15].

• Equivariant diffusion — Encoding the world as a point cloud enforces physical consistency, meaning rotations or shifts in input data create predictable, correct transformations in the output [08:35].

• Image-to-sphere — This method projects visual data onto a sphere to maintain coordinate integrity, allowing the system to process camera inputs while preserving 3D orientation [29:27].

• 3D ray representation — Converting image patches into vectors pointing from the camera helps the system map visual inputs into 3D space, which facilitates the combination of multiple camera views [36:36].

• Viewpoint robustness — Using data augmentation that simulates camera spinning forces the model to focus on local visual features rather than global context, improving performance across varied viewpoints [46:49].

• Scaling laws — These structured models shift performance curves, allowing robots to achieve high accuracy with less total training data than purely data-driven models [52:07].

• How does the speaker define "equivariance" in the context of robot policy learning?

• Why does the speaker argue that incorporating structure is more effective than relying on "big data" models?