Graphics + Vision Research Seminars occur on Mondays in Gates 114, from 2:55-4:10 p.m.
The Graphics + Vision Research Seminar discusses recent research in the areas of computer graphics and computer vision. The goal is to foster technical discussions and collaboration among the Cornell graphics and vision research community.
Fall Schedule TBD
Stay in the loop.
Date: September 14, 2026
Speaker: Junho Kim
Title: Finding Analogies Everywhere: From 3D Scenes to Touch to Historical Maps
Abstract: Analogies are ubiquitous. From the now-obsolete J.J. Thomson’s “plum pudding” model of the atom to the “greenhouse gases” warming our planet, analogies allow for conveying complex concepts by borrowing illustrations from simpler, well-known entities. In this talk, I will introduce recent works that find analogies from visual inputs and 3D scenes, which consequently enable transferring complex concepts such as motion trajectories, scene arrangements, and tactile information. Specifically, I will illustrate how 2D/3D foundation model features can be leveraged to find alignments between different domains, and how the alignments can be further used to smoothly map concepts in one domain (e.g., motion or touch) to another. At the end of my talk, I will share progress in my recent work on understanding historical maps, where we aim to find alignments between historical maps and modern maps to model how cities change over time.
Speaker: Chuanruo Ning
Title: Proxy Policy Steering
Abstract: Generalist robot policies carry broad manipulation priors from large-scale data, but specializing them to a new task remains the deployment bottleneck. This requires eliciting task-specific behavior from limited demonstrations without degrading their broad capabilities. We introduce Proxy Policy Steering (PPS), an inference-time adaptation method that resolves this challenge by training two lightweight proxy policies whose calibrated velocity-space difference steers the frozen base sampler. A reference proxy models the frozen base's behavior on target-task observations, and a task proxy, initialized from the reference, captures how this behavior changes under task supervision. Their difference forms a calibrated velocity-space residual that steers the frozen base sampler at every denoising step. We identify the conditions under which this residual isolates the change induced by task supervision, and validate them empirically. Because the base is never directly modified, its broad capabilities remain available at inference, including behaviors such as recovery from failure that the demonstrations themselves do not exercise. Adaptation requires only forward velocity predictions from the base, making PPS lightweight to train and applicable even without access to the base's parameters. On 8 real-world and 4 simulation manipulation tasks, PPS lifts the state-of-the-art pi 0.5 base policy by 53% absolute success rate on average, with zero-to-one gains on tasks the base never solves, while preserving the base's broad capabilities. PPS outperforms LoRA fine-tuning, from-scratch specialists, residual policies, and prior inference-time steering methods.