Closed-loop visuomotor control with wrist cameras is widely adopted and powerful, but leveraging full 3D symmetry from only RGB is still hard. Introducing our #NeurIPS2025 Spotlight paper, Image-to-Sphere Policy (ISP), for equivariant policy learning from eye-in-hand RGB images.

Do you want your network to reason about 3D symmetries present in 2D images? Consider using our induction layer, which is theoretically justified and outperforms existing approaches. Ask Owen Howell about it at #NeurIPS2023 today, Wed. Dec 13 5pm, Great Hall & Hall B1+B2 (lvl1)

We also show our model can learn complex distributions over 3D rotations, by parametrizing the distribution in Fourier domain of SO(3). Without any modifications, our method handles objects with discrete or continuous symmetries.

We introduce a mapping to project learned features from the image to the sphere. Features on the sphere are then processed with SO(3) equivariant layers. By encoding SO(3) symmetry of the problem into parts of the network, we achieve higher accuracy than existing baselines,

Our work is titled “Image to Sphere: Learning Equivariant Features for Efficient Pose Prediction”. PDF: https://t.co/BMI1Xmkl6U. Webpage: https://t.co/3vdHuPrNL0. Collaboration with @BizaOndrej, @RobotPlatt and @RobinSFWalters.

How can we efficiently model uncertainty over 3D object orientation given a single image? Our #ICLR23 work shows how to leverage SO(3) symmetry to better learn complex rotation distributions. My talk is on Tues. (May 2nd) in Oral3 Track2 (AD12), & Poster Session 3

This was joint work with @BizaOndrej and @RobotPlatt. Find the full paper and code at

I’m presenting our paper “Graph-Structured Policy Learning for Multi-Goal Manipulation Tasks” at #IROS2022 tomorrow. Check it out in Room C-1 at 11am