NYC Physical AI Meetup Connects Robot Learning and Vision

The NYC Physical AI, ML, and Computer Vision Meetup will bring five announced speakers to Fractal in Brooklyn on November 13, 2026. Co-presented by Nebius and Voxel51, the gathering runs from 5:30 to 8:30 PM EST and connects robot learning, computer vision, video retrieval and model representation in one evening of lightning talks and networking.

The common thread is the work of turning visual experience into something a machine can use. A demonstration has to become training material, a recording has to become searchable, and an engineering team has to understand why a system failed before choosing its next fix. For technical founders and operators, the agenda offers a chance to follow those decisions across specialties while the products and research are still being built.

Five speakers, several versions of the same handoff

The official Voxel51 agenda lists Dan Gural of Voxel51, Dhruv Diddi of Nebius, Irmak Guzey of New York University, Yash Gugale of TwelveLabs and Yunied Puig de Dios, whose affiliation is listed as Independent. Their subjects range from practical vision development and physical AI to dexterity, video search and OrBIT compression. That breadth gives the evening its interest: an engineering decision that looks local can become somebody else's constraint a few steps downstream.

A researcher working on a robot hand and an engineer helping people search video may have very different definitions of success. Both still need a useful representation of what happened, a way to inspect mistakes and some basis for deciding whether an improvement survives outside the chosen example. The productive conversation sits in those specifics, where a failure can be described clearly enough for someone from another discipline to recognize part of it.

The meetup is designed around talks followed by time to connect, with food and drinks announced on the registration page. Engineers, researchers and practitioners have an obvious fit with the technical program. A founder or product leader can make the same room useful by bringing a precise integration question, rather than expecting the agenda to deliver a finished answer to a broad AI strategy problem.

Robot learning makes the data problem tangible

Irmak Guzey's talk addresses the gap between human demonstrations and dexterous robot hands. Guzey's research focuses on robot learning and manipulation, while the primary HuDOR research paper examines learning from human video despite differences between human and robot bodies and visual observations. The difficulty has a wonderfully stubborn physical quality: seeing a person complete a task does not give a machine that person's fingers, viewpoint or control over an object.

For an operator evaluating a robotics project, that raises a practical question about the next unit of effort. Should the team collect different examples, change how the task is represented, or revisit how performance is evaluated? The answer affects whose work comes next. A discussion about data collection can therefore become a discussion about engineering ownership, even before anyone starts talking about a commercial rollout.

Dan Gural's computer-vision session and Dhruv Diddi's physical-AI introduction offer adjacent ways into that conversation. Voxel51 describes its platform around finding, visualizing, curating and evaluating multimodal data; Nebius supplies AI cloud infrastructure for training and inference. Those company descriptions help explain the mix of perspectives. The value for attendees will come from asking how the pieces connect to a particular task, with its own observations, errors and operating constraints.

Video search brings the question back to the user

Yash Gugale's announced session considers how a natural-language query can locate material across hours of footage. TwelveLabs builds video-understanding and retrieval technology, and Gugale's own biography describes a customer-facing engineering role at the company. That position matters to the discussion because a useful search result has to meet somebody's actual request, with all the ambiguity that arrives when a person describes what they remember seeing.

A retrieved frame can look convincing while leaving out the sequence that gave it meaning. An operator considering video search should therefore arrive with questions about how to inspect results, define a useful match and handle an answer that is visually plausible but operationally wrong. These are editorial questions prompted by the topic, not reported conclusions from a talk that has yet to happen. They give product and engineering teams a concrete conversation to continue after the presentation.

Yunied Puig de Dios's OrBIT session adds representation efficiency to the evening. Compression introduces another set of choices about what information a system retains and how that representation can be used later. Taken together, the agenda gives attendees a way to compare the cost of carrying visual information through a workflow with the cost of losing something important along the way, whether the eventual output is a search result or a robot's movement.

What to bring to the Brooklyn conversation

The venue is Fractal, also labeled Fractal Tech on Luma, at 111 Conselyea St, Fl 2, Brooklyn, NY 11211. Voxel51 and Fractal NYC appear as hosts on the registration page, while Nebius and Voxel51 are named as co-presenters. Readers should use the official registration page for current access details; the available source material does not establish a ticket price, livestream, moderator or guaranteed admission.

The nearby operating conversation also runs through DevCuration's coverage of the NYC Industrial Tech Breakfast and enterprise AI data ownership. Those are separate gatherings, but they offer readers other entry points into the decisions around putting AI to work. No shared participant or organizer relationship is implied.

For November 13, bring one example that another builder could examine: a recording, a failed grasp, a search request or an evaluation decision the team keeps revisiting. The interesting exchange can begin when the person across the table asks for the detail your own team has stopped explaining. In a room spanning robot hands, video retrieval and data tooling, that detail may send the next conversation toward a different engineer than the one you expected to need.