Silicon Valley Physical AI Workshop and Meetup - October 27, 2026

Oct 27, 2026

5:30 PM - 8:30 PM PDT

Silicon Valley AI Hub, 135 Constitution Drive, Menlo Park, CA 94025

Speakers

About this event

Join us at Silicon Valley AI Hub on October 27th for the Silicon Valley Physical AI Workshop and Meetup, co-presented by Nebius and Voxel51.

The evening opens with networking, food, and drinks, followed by a hands-on workshop on building physical AI applications with Voxel51, from curating robotics and sensor datasets to evaluating vision models on real-world data. Networking continues after the workshop.

A laptop is required to participate in the hands-on workshop - please bring one. Space is limited, so register early.

A company, university, or government email address is required to take part in the hands-on workshop. Personal and disposable email accounts (Gmail, Yahoo, Outlook, and similar) cannot be used for the workshop environment. This is a blanket policy of the workshop platform to prevent fraud and abuse and to meet compliance requirements. If you need an exception, contact the hosts before the event.

Schedule

Hosts

Teaching robots from video: no robot lab required

I build tools that turn a single video or sentence into physics-checked robot demonstrations. This talk walks through Player Two, which turns webcam video into gated LeRobot demonstrations in browser MuJoCo, and Understudy, which uses generated video the same way.

A SmolVLA policy trained on Understudy demonstrations scored 43 out of 50 in simulation, against 44 out of 50 for perfect scripted demonstrations. I will show a live demo of the pipeline and the checks that reject bad demonstrations. Attendees leave with a recipe for collecting robot training data without a robot lab.

Testing AI Agents Beyond the Completion Message

An AI agent can report success while a tool action remains unresolved, has been repeated, or was never authorized. Drawing on Sajal Garg's automated LLM evaluation work at Microsoft AI, this talk shows how to check observable actions and service state alongside the quality of generated responses. A non-confidential illustrative task demonstrates why interrupted execution, lost acknowledgements and resettable test state belong in an agent evaluation workflow. The session separates task failures from infrastructure and evaluator failures, then compares current and candidate versions using the same cases and repeated runs. Attendees leave with concrete action/effect assertions and a release checklist that keeps critical execution failures separate from aggregate quality scores.

Agenda

5:30 to 6:00 PM

Doors open, networking, food and drinks

6:00 to 7:00 PM

Workshop

7:00 to 7:30 PM

Lightning talks

7:30 to 8:30 PM

Networking and close

Workshop Agenda

This hands-on session uses DROID, a real-world robotics dataset loaded into Voxel51 as a native multimodal MCAP recording, and YOLO11n, fine-tuned live during the session.

Welcome + framing: from raw robot logs to a trained detector
Explore a real DROID robotics recording in Voxel51's native multimodal MCAP viewer: camera, proprioception, and language on one synced timeline, no ROS install required
Curate: extract and browse frames from the recording, filter and deduplicate
Compute embeddings on the curated frames; explore via similarity search and embeddings visualization (via Nebius Serverless AI Jobs)
Auto-label: open-vocabulary detection to generate bounding boxes for the robot gripper and target objects
Train: fine-tune a YOLO11n detector on the auto-labeled frames (via Nebius Serverless AI Jobs)
Evaluate results and close the loop: view predictions back on the original MCAP timeline
What else Nebius offers: Token Factory walkthrough — chat/vision models, fine-tuning, credits